I decided to take the plunge and get myself four sparks at a decent price (and bought the QSFP cables from AliExpress because they are literally 1/2 the price of Amazon), even knowing Apple was going to release new hardware and there's probably a spark 2 on the horizon. It looks like this is going to be a decent fit for what I need. I've been experimenting with a two-node DS4 and it's _good_ at some tasks, but it really just spins its wheels when it hits the limit of what it can reason through.
I can offload mundane/basic tasks to DS4 on two sparks, but I've been pushing it harder on some novel work and it just can't run on its own at all beyond a certain complexity level.
I would love to see an Opus-4.8-level local model but TBH I just haven't got there yet. The models I've tried so far _are_ good but they aren't able to solve tough technical challenges, regardless of harness/prompting/etc.
Qwen 3.8 27B is around Opus 4.8 level of capability on the Agentic Intelligence Index (52 vs 57). In my testing the locally hosted Qwen is good enough that looking at a given piece of work output I couldn't tell you which model was behind it.
I will give it a try, but from the benchmarks it never exceeds the DS4 flash benchmarks by significant margin and And I feel that the throughput that you will get on those machines or what I'm getting with my local hosted flash will be so much worse that it's not worth it.
Compared to pricing from 3 years ago, it's insane.
Compared to 1-2 years worth of LLM tokens for a full-time software engineer, $12K for on-prem private LLM inference starts looking reasonable, particularly if privacy is an important consideration.
The Sparks admittedly are kind of anemic: 273GB/sec is the same bandwidth as a midrange 4060, although (depending on how you configure things) you can effectively have much greater bandwidth by connecting them.
Also, it is going to be a long time until RAM+GPU prices return to what we used to call "normal." If ever. I am not endorsing the current state of affairs and I am not saying you wrong to find it insane, but it is definitely the new reality.
Compare to the cost of professional-grade tools in other trades and craft hobbies.
Sure, $4000 can be a lot of if you're a casual hobbyist or are struggle to meet everyday lifestyle costs, but it's definitely not "insane" if this is the trade you make your living from or if you've established a lifestyle that affords disposable income for your hobbies.
And for some people, $4000 for a device you have complete control over and can repurpose and tinker with to your own needs and curiosities is a much much more justifiable expense than a $200/mo rental for some narrow-access tool that somebody else controls.
The other reason is that it would likely take years to spend $4000 (plus the real cost of electricity) worth of tokens on a 3rd-party provider that's running a similar limited, DS Flash type model. By that time, the hardware will be obsolete, assuming it's still operational.
> it would likely take years to spend $4000 (plus the real cost of electricity)
Since that cluster only yields 20-30 tok/s on that size of model, at least a decade before the hardware breaks-even with current token costs, and that's not counting electricity. Assuming continued downward pressure on token prices, and the cost of electricity, it never pays for itself.
I am pretty confident that given a $200 subscription on any of the big labs, you're getting $4000-$8000 per month in subsidized tokens... do what you wan't with your dough... and I too have a spark that I got really early (October 2025), but no, economically it does not compare to what's runnable locally in terms of quality from the frontier models. Economically, it looks like for as long as there are subscriber plans, you're better off renting.
Before getting the spark, I was just using a google colab account, their $49 dollar plan allows you access to h100's and I can run qwen there in a Jupyter notebook... and if I really need that web front end I can just use cloudeflair/tailscale/the local ssh client to reverse tunnel it.
The cloud stuff is definitely a much better economic value, but I would argue:
1. You learn a lot more running this stuff yourself (especially since you can poke at its internals if you're interested or watch the reasoning chain.) Just being a consumer of this stuff doesn't really teach you much about it other than model & harness specific tricks that become obsolete pretty quickly. (IE, your Claude.md from 6 months ago probably needs a rewrite). Which is fine, I don't think you're going to be "left behind" if you're not a hardcore AI enthusiast or anything (I'm not), but as a guy that's always been interested in computer science I want to see how it ticks.
2. You can't really depend on this subsidization lasting forever IMO. I know the financials thing has been beaten to death but I guess I'm in the camp that it's good to be in control of your tools so that you can go elsewhere if the economics change.
I like to check in with ccusage pretty frequently, and honestly like if I were paying API prices for Claude I'd probably be paying thousands a month.
I am also not sure I would choose to use the cheap and easy to run at home model, given a choice. The marketing copy says this is a frontier model, but it's not. Sol and Mythos are the frontier right now. GLM 5.3 Flash simply isn't. I'd rather use the frontier model as they waste less of my time than even Opus.
I mean, I have the same machine and the pricing is only what it is because it has that 1TB nVME in it instead of larger. nVME prices are insane and have been for months.
Reality is on a single spark I'm constantly running out of room and it being an odd size M.2 slot it's a pain to upgrade. I'm setting up a NAS over RDMA via ConnectX though, that's fun.
I already have a recycled QNAP-now-TrueNAS that has 10G connections into the fleet so the 1TB doesn't bother me at all. I did some rough math and I don't think I'll ever need to load weights off NFS for what I'm doing so far, but the capability is there.
> I would love to see an Opus-4.8-level local model but TBH I just haven't got there yet. The models I've tried so far _are_ good but they aren't able to solve tough technical challenges, regardless of harness/prompting/etc.
Agree. It doesn’t even have to be local, using models in this size class through OpenRouter will reveal their limits if you work side by side with Opus level models regularly.
There are a lot of social media posts about people cancelling their Anthropic or ChatGPT subscriptions after installing a local LLM. I’ve used local LLMs a lot and I spend a lot of time with frontier models and the difference is still huge. As far as I can tell, the social media posts about local LLMs replacing frontier models are either wishful thinking, engagement bait, or people who must be working on much simpler projects with a much higher tolerance for slop than I have.
Over the last couple years I’ve had to learn sales and understand the thought process behind this better, and I think I’m beginning to understand it
The psychology is that most people aren’t really trying to optimize for productivity (even most people who think they are) on an ROI basis, because their compensation is too decoupled from their actual raw output, and more closely coupled to how differentiated their marginal contribution is to peers. They’re much more incentivized to spend their personal/work time optimizing for being more skilled or acquiring some kind of competitive advantage relative to baseline.
Most people don’t consciously run the numbers of “I get paid $X/hr to add $Y of value” or model pay at work as something with variable inputs (eg something that can be increased with high performance), so it makes sense to them to spend 20 hours of time to save $100 or to make themselves 5% less efficient to take home 0.5% more or avoid doing something they don’t want to start doing.
NOT saying this always happens or that they’re stupid for doing so. I didn’t even realize how much I had been doing it myself until I started recognizing it, and shifted to having my own comp/performance fully aligned with the company’s P/L.
It actually makes a lot of sense IF you can accurately estimate incremental upside (which is much harder and more diffuse than modeling downside if you’re salaried a employee) or if the upfront skill/knowledge investment that looks like bikeshedding pays off in the long run.
These are great points. It's a little off topic but what you bring up is why i advise new grads to spend the first couple years of their career in small eat-what-you-kill companies. I think software devs who start out in large companies get this distorted view that their twice a month direct deposit is just magic and comes from the ether no matter what they do. The whole industry would be better off if everyone started out in a "you don't deliver, you don't eat" company and grew from there.
To be fair, there is no 3 turns that I don't have to jump in into what Opus 5 is doing. There is either some regression or my prompting skills are so much worse now. Flash is not perfect and honestly some things depend on how big context do you keep. So I'm keeping like a really short context with my flash, but it works okay, even though it has a tendency to overthink, and yeah, I run it always in max effort mode.
If you used the bare API pricing, 1M tokens @ 30% input/70% output/50% cached, you'd pay $0.05805. Even with four discounted sparks, how much are you paying for the same tokens/distribution?
There's soooo much by way of experiments, explorations, tinkering, and even projects that you can't possibly pursue through a some SaaS API.
The more reasonable comparison is against rented GPU's, while looking at tradeoffs in latency and upload/download/storage/instance management overhead.
Buying hardware for local models is meeting a wholly different need than buying tokens through OpenRouter or whatever.
For me it's entirely because I have a bunch of projects with my own personal data that would be tough to do with openrouter/claude or any other cloud.
For example, I have a small posix-shell-based LLM harness that can SSH into my NAS and run organization tasks using the local DS4Flash that I have right now. It's already been a massive help for me to keep me organized, and that's just 2x DGX Spark's worth of compute.
It cuts both ways. A GPU in your basement is a depreciating asset with fixed computing power and consumes electricity. Switching model providers is trivial.
All decades prior and up to about a year ago, I would have agreed with you. My Framework Desktop, however has appreciated in value by 75% since I bought it. Will it stay there for a long time? Probably not. But it shows that there are no hard and fast rules about things anymore.
Chinese labs are so used to manipulating benchmarks to try to flatter inferior models that when they finally have one that's really pretty good I think the official announcement here undersells it.
That's pretty solid. Smarter and cheaper than Luna xhigh, not as smart but less expensive than Luna max. Smashes deepseek v4 flash, and even worse it matches v4 pro at a tiny fraction the cost. Roughly equivalent to sol medium, at a fraction the cost.
They should've just lead with real, up to date data, because it's good, not the silly old tactics like comparing to Opus 4.8 when 5.0 is out in many of their charts.
Maybe others have found otherwise, but I find the benchmarks drastically different to real world "feel" of a model, even within the same harness. I'm not sure if this just reflects personal interaction styles, or if it is indicative of benchmaxxing or unrealistic automated benchmarking methodology.
Opus 5 consistently comes at or near the top, but outputs constant unreadable jibberish. Meanwhile GPT 5.6 Luna medium tends to be rated pretty poor on agentic tasks compared to the Chinese lab open models, but I find the latter much more likely to lose track of their own behaviour during a long-horizon task or get stuck in a doom loop.
(This isn't a comment on GLM-5.3 Flash as I've not used it!)
I've been using 5.3 since they initially announced it and my gut feel is that it's not as good as 5.2 for agentic tasks. I'm still using it - I don't think it's bad. I'm just not convinced it's better.
Even Artificial Analysis has Opus 5 better than Fable in their aggregated "Intelligence Index" which combines 9 benchmarks. Opus 5 is heavily benchmaxxed.
> They should've just lead with real, up to date data, because it's good, not the silly old tactics like comparing to Opus 4.8 when 5.0 is out in many of their charts
It's what people know. Opus is just the common target.
> Smarter and cheaper than Luna xhigh, not as smart but less expensive than Luna max. Smashes deepseek v4 flash
The problem with this and DeepSWE is it goes for a very specific profile. I'm not convinced DeepSWE is any accurate in actual work. It's surely a different signal (compared to some that allow cheating) but it has its own issues, e.g. weak harness.
Luna is great at following instructions but bad instructions or anything not covered = death.
Deepseek is more analytical. Good for bug tracking.
GLM is a better all rounder in some ways. Better at creativity.
Broad and perpetual license over inputs and outputs, and even your name and profile picture.
Vague prohibitions on whatever may harm Z.ai’s "interests" or even the "national interests" of any country.
Vague prohibitions on "disturbing" or "inappropriate" content, whatever that is.
Vague prohibitions on discussing Z.ai, even my posting this comment violates it.
Can ban you if you, in the "sole and absolute opinion" of Z.ai, have violated these broad terms, and if you paid for the discounted yearly plan kiss your money goodbye.
Nearly every TOS I've ever read has a "We can ban you for any reason, or no reason, are under no obligation to disclose any reason." line somewhere in it.
HN's for example
> We reserve the right, at our sole discretion, to change or modify portions of these Terms of Use at any time.
> You acknowledge that Y Combinator may establish general practices and limits concerning use of the Site,
> You further acknowledge that Y Combinator reserves the right to change these general practices and limits at any time, in its sole discretion, with or without notice.
> Y Combinator reserves the right to investigate and take appropriate legal action against anyone who, in Y Combinator’s sole discretion, violates this provision, including without limitation, removing the offending content from the Site, suspending or terminating the account of such violators and reporting you to the law enforcement authorities.
None of the major LLM chat providers (ChatGPT, Claude and Gemini, and I just confirmed this) claim rights over your input.
They also don't claim rights over your output, but because of how copyright law might apply, they explicitly assign all the rights to the generated output.
Not just that but, even if they wanted to claim ownership of the output, courts in the US have deemed that copyright cannot be assigned to machine-generated output.
So yeah, happy to take a look at a counter-example if you have one (aside from GLM 5.3, obv.).
> In choosing to submit, create, generate, record, post, or display Inputs on or through the Service, you grant an irrevocable, perpetual, transferable, sublicensable, royalty-free, and worldwide right to SpaceXAI to use, copy, store, modify, process, adapt, transmit, distribute, reproduce, publish, upload, download, display in public forums, list information regarding, make derivative works of, and distribute such Content, including anything referenced therein, in any and all media or distribution methods now known or later developed, for any purpose, and to aggregate your User Content and derivative works thereof for any purpose, including but not limited to: (i) maintain and provide the Service; (ii) improve our products and the Service and for our other business purposes, such as data analysis, customer and market research, developing new products or features, or identifying or displaying usage or User Content trends; and (iii) perform such other actions to enforce these Terms, comply with our Privacy Policy, comply with applicable law or governmental, court, and law enforcement requests or requirements or keep our Service safe.
> To the extent the User Content includes a person’s image, likeness, voice, or other similar attributes, you grant SpaceXAI the same rights to use those attributes as part of the User Content as described above. You represent and warrant that you have obtained all rights, licenses, notices, permissions, and consents necessary for SpaceXAI to use that User Content.
- Grok - where I absolutely have 0 trust in X.ai's interst in "pushing humanity forward".
- OpenAI and Anthropic - which seem to try to be building the biggest moat they can by pushing to ban open models. And at the same time want to be an Arbiter of what level of intelligence I can use.
- Google and Meta - I don't need to talk about the practices of these companies.
Yes, the terms of service aren't great. But the alternatives aren't great either. I don't believe that a future which OpenAI and Anthropic are pushing for has my best interest in mind.
All the American companies you mentioned still follow American law and regulation. Skirting that blatantly has big consequences.
Chinese companies do not follow American laws and there are absolutely no consequences for violating it.
Moreover, the average American is not even aware of exactly what the legal/judicial environment is like in China. If your code and data is stolen, you can't fly to China and demand justice in the courts.
Give it a couple days, and there will be plenty of other inference companies hosting it. Don't like z.ai's TOS? Use the model on a provider with TOS that you agree with.
I blocked Z.ai as soon as they were loading 10 different external providers including Alibaba who was just proven to execute silent sound fingerprinting mechanisms.
With tiny models surpassing huge, 6 months old models on benchmarks, does anybody have some smart words to share on how these still "feel" different?
Artificial Analysis ranks GPT 5.6 Luna similar to GPT 5.4, but that never matches my real world experience. AA seems to do a good job making a single number as representative as possible but there is still so much benchmarks don't communicate.
I agree. They are definitely good - no issues with instruction following for example - but they miss the "intelligence" larger models have.
For implementation tasks, where I have the problem already defined and researched, or just simple task, I'd definitely use something like Luna xhigh or max. If the task is vague, or involves planning, I'd rather use Sol medium, even though it's theoretically worse on benchmarks.
Will we need all the data centers being built or will improvements in software and hardware allow the majority of AI workloads to run locally or in the cloud but way more efficiently than was projected when all the plans were laid out?
Like were executive at Google and AWS and Microsoft expecting this kind of performance from models smaller than what openai/anthropic have been doing? Are we really in a "compute desert"?
I don't see a situation where subscription payers move outside American LLMs (chatgpt, claude, gemini)
And I don't see a situation where serious API payers are OK with handing the Chinese state all their data. Like manufactures of decades past did and learned a hard, even existential, lesson for it. The state mantra has been "Collect and Copy" for a long time now, tech just hasn't had that moment to experience it yet.
So that leaves local hosting/leasing, but one of those has totally non-practical economics and the other doesn't have enough compute to meet any kind of real demand.
I also have yet to meet a single person who isn't neck-deep in the tech space mention a Chinese LLM. It's 100% the big American three.
If anything it's custom chips from the labs that threatens Nvidia.
Google probably serves more tokens then OAI and Anthropic combined, even if many of those tokens aren't from explicit gemini requests, but from AI overviews and other service integrations.
xAI is already selling spare compute, and basically exists just to gas spacex's perceived valuation.
This is the takeaway here: That's how they have been serving it at scale as Ox-Alpha. This is a definitional moment.-
Further quote:
"Compared with our initial baseline on the same hardware, we achieved a 3× improvement in end-to-end serving performance, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs. This demonstrates that Chinese chips can support frontier-model inference efficiently and economically at scale."
Another self-inflicted own courtesy of US government policy.
While I think China would always get to hardware self-sufficiency eventually, all export controls have done is (1) accelerate China's development, and (2) divert revenue that would've otherwise gone to NVIDIA/AMD/etc instead.
Long term it's irrelevant. The only relevant thing is that there's lots of money in chips that can do high performance inference. You see all kinds of competitor products in development or already on the market even here in the US where there are no such restrictions. Cerebras comes to mind. It's natural and expected that eventually Nvidia will either have to keep way ahead or competition will catch up with specialized products.
That doesn't mean by any stretch of the imagination Nvidia will disappear. But the entire stock market valuation, not just tech, has had me scratching my head for a while.
Revoked or not, just ever having those controls signals to the Chinese ecosystem that you're not necessarily a reliable supplier (Would you trust US export policy to remain stable for the next ~decade given the state of US politic?) and to the Chinese government just how strategically important you see these components.
This isn't the kind of thing you can hash out in public and go back and forth on. Once you put it out there, the other party will take steps to make sure they don't have to rely on us in the long run.
The export controls were not revoked, only reduced, and not before, but after China refused to buy low performing chips. Top gear was and is still sanctioned, as is any EUVL equipment.
And to add to the above: by building their own supply chain for chips, China is helping the unprivileged, those who can't front-run the market with long-term contracts. If China wasn't producing their own chips, the prices for us would be even higher.
Similar to the war-pricing of oil, China's reduction of imports is actually helping to keep our inflation from going even higher.
It's also in this very announcement, in the first paragraph:
> Before release, we tested GLM-5.3-Flash anonymously as ox-alpha on OpenCode and OpenRouter to gather user feedback. It quickly became the most popular model of the week — with all of this traffic served on Chinese AI chips.
>> "They are already there on open weight models and Jensen knows that it is only a matter of time until China catches up with GPUs or other AI accelerators."
It is also why Nvidia becoming a bank for other AI companies who are unable to find VCs to fund them isn't really a good thing and that is bearish.
Is the actual Z.AI ecosystem good enough to replace the main drivers like Codex and Claude? Because it looks like Z Code is just a Codex fork. Just like the Kimi Code one is.
What irks me about this is that the harnesses seem to be just an afterthought here.
Don't get me wrong, I love messing around with installing Pi, getting it hooked up with OpenRouter, and just trying all kinds of different stuff, local models, etc... but when it comes to literally just setting up a productivity environment and trusting my entire machine with it, I just run Codex.
I have heard from anecdotes where people have indeed replaced their main drivers with DeepSek V4 Flash or GLM and state that "it's almost as good as... [claude/gpt]" but I never hear anyone say "yeah, this is the model/harness that I now run on my machine and don't mess with it"
> 320B total parameters and just 18B active parameters
This is pretty hefty for a "flash" model, even a 256 GB setup is insufficient at q4 - and q4 is already the worst-but-still-acceptable quant in my experience. The benchmarks look great, especially since GLM tends to be more honest than the average Chinese lab, but you’ll need to splurge to run it at home.
@edit: so many releases that I forgot to math. This fits just fine in q4, realistically the minimal hardware would be 192gb - so blazing fast on double rtx 6000 pro and usable on 256gb unified memory. You could even go with 5bit quant on 256gb.
> it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks.
From a biased source, but would be big if true.
I've had great results with GLM 5.2.
From their subscription page, the smallest plan gives you about 97M tokens weekly for 5.3 but 292M for 5.3 Flash. Not exactly 10x the limit.
│ https://openrouter.ai/api/v1/chat/completions model: stealth/ox-alpha auth: OPENROUTER_API_KEY status: 404 Not Found response: {"error":{"message":"Thank you for participating in the Stealth Ox Alpha testing period. This model was ZAI's GLM-5.3 Flash.
When reading this type of announcements, always have keen eyes on graphs.
e.g. "Agent Coding Performance by Effort Level" cuts Y-axis from 0~20.
- This makes it as if GLM-5.3-Flash made a bigger jump than it claimed as the Y-axis does not increase much (stupid trick used in biz reports)
I did mention that ox was working ok for me, and having an open-weight comparable to close to SOTA makes it very compelling for me to try it out locally (well, only if I got more VRAM)
Well, Luna debuted with 5x higher pricing than is currently available. With the pace of recent development these models might not be relevant by Thanksgiving.
Of course. Pricing is always changing, but typically it goes down over time, not up. So, if you're showing artificially low pricing from the start based on a teaser rate, IMO, you shouldn't be using that to show where you appear on a frontier graph. Place yourself on the graph based on your expected long-term pricing. Then, over time, adjust your position based on your standard rate, whatever that might be. Games are always being played for things like this, but this seems excessive.
I really wish GLM models had vision capabilities. I've worked around that in the past to use a vision MCP in my harness that GLM can call. It is not the same, but it allows the model to query images.
At the current 50%-off GLM-5.3-Flash price ($0.075/M input, $0.25/M output; cached input $0.015/M), surprisingly, roughly $400–900/month would buy token throughput comparable to fully exhausting Claude Max 20×
They're obviously in a pickle, nobody is going to continue to pay $15-50 a mm tokens here soon. There's a reason OpenAI stopped training large models last week, and it's not because of "saftey" or "alignment" they know these gigantic models are not worth the squeeze.
If we fast forward say 5 years, I don't see how we don't end up in world where people (and enterprises) are more savvy with how they use LLMs. Meaning, more models, smaller models, weirder models, more specialized models, etc. And all of it running on a variety of hardware (edge devices, personal computers, on-demand cloud compute).
I don't see how NVIDIA can keep their spot as belle of the ball. If LLMs and friends are truly to become as useful and ubiquitous as everyone thinks they will, then commoditization is the only option.
We need to figure out what the real pricing is for a going concern. Right now, everyone is subsidizing and discounting to grow (or maintain) market share. The big question is whether the steady state, market derived inference pricing is above or below what we’re seeing today. I honestly don’t know. Anthropic had said that inference is profitable, but they’re clearly not yet profitable overall with training and buildouts still happening.
Well the big problem with china is that they do not respect international law when it comes to technology theft. But that argument is very weak when it appears that a lot of what they do is out in the open for anyone to replicate.
I find GLM's idea of fast/flash is not really competitive with the speed DS4 Flash has, and it's hard to see them as being in the same segment for that reason.
All I can say is that even if it is, I was almost glad to go back to using DS4 Flash. Because 0XAlpha was just so friggin slow to complete a task because of the level of circular reasoning that it would go over and over into, sometimes even returning no output. If I just wanted something done I would switch from a free model to a paid one which is crazy.
Possibly influenced by that, but I believe that is a different issue. I meant the way it processed a request. It went into so many more loops of thinking.
> To overcome the relatively limited compute and memory capacity of individual chips, we built a dedicated inference engine for this architecture on top of SGLang. Notably, this effort was accelerated by our GLM-5.3-powered infrastructure agent, which assisted engineers in developing and optimizing kernels, diagnosing performance bottlenecks, and improving the serving stack — creating a feedback loop in which the model helped optimize the system serving the model itself.
> (...) Compared with our initial baseline on the same hardware, we achieved a 3× improvement in end-to-end serving performance, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs. This demonstrates that Chinese chips can support frontier-model inference efficiently and economically at scale.
It might be one of the most actually practical tasks that AI might've done because the compounding effects of it and also its implications are/feels so immense. It feels as if Nvidia might be in a slight turbulence from it.
I didn't accept a single edit from this model over the entire week, just saying. I do not understand how it's being benchmarked on par with Sol and other larger models.
Weights on HF here: https://huggingface.co/zai-org/GLM-5.3-Flash
I decided to take the plunge and get myself four sparks at a decent price (and bought the QSFP cables from AliExpress because they are literally 1/2 the price of Amazon), even knowing Apple was going to release new hardware and there's probably a spark 2 on the horizon. It looks like this is going to be a decent fit for what I need. I've been experimenting with a two-node DS4 and it's _good_ at some tasks, but it really just spins its wheels when it hits the limit of what it can reason through.
I can offload mundane/basic tasks to DS4 on two sparks, but I've been pushing it harder on some novel work and it just can't run on its own at all beyond a certain complexity level.
I would love to see an Opus-4.8-level local model but TBH I just haven't got there yet. The models I've tried so far _are_ good but they aren't able to solve tough technical challenges, regardless of harness/prompting/etc.
Qwen 3.8 27B is around Opus 4.8 level of capability on the Agentic Intelligence Index (52 vs 57). In my testing the locally hosted Qwen is good enough that looking at a given piece of work output I couldn't tell you which model was behind it.
https://artificialanalysis.ai/models/qwen3-8-27b?models=gpt-...
I will give it a try, but from the benchmarks it never exceeds the DS4 flash benchmarks by significant margin and And I feel that the throughput that you will get on those machines or what I'm getting with my local hosted flash will be so much worse that it's not worth it.
Hopefully you also bought a switch
> get myself four sparks at a decent price
Wow, if you don't mind me asking. How and where?
I bought 4x Asus GX10 with the 1TB option. I don't understand why, but it's the only model in the whole lineup that isn't priced insanely.
They were briefly on sale with a $200-off coupon, but they show up on warehouse deals from time-to-time as well.
> it's the only model in the whole lineup that isn't priced insanely
$4,000 isn't priced insanely? ye gads
Compared to pricing from 3 years ago, it's insane.
Compared to 1-2 years worth of LLM tokens for a full-time software engineer, $12K for on-prem private LLM inference starts looking reasonable, particularly if privacy is an important consideration.
The Sparks admittedly are kind of anemic: 273GB/sec is the same bandwidth as a midrange 4060, although (depending on how you configure things) you can effectively have much greater bandwidth by connecting them.
Also, it is going to be a long time until RAM+GPU prices return to what we used to call "normal." If ever. I am not endorsing the current state of affairs and I am not saying you wrong to find it insane, but it is definitely the new reality.
Compare to the cost of professional-grade tools in other trades and craft hobbies.
Sure, $4000 can be a lot of if you're a casual hobbyist or are struggle to meet everyday lifestyle costs, but it's definitely not "insane" if this is the trade you make your living from or if you've established a lifestyle that affords disposable income for your hobbies.
And for some people, $4000 for a device you have complete control over and can repurpose and tinker with to your own needs and curiosities is a much much more justifiable expense than a $200/mo rental for some narrow-access tool that somebody else controls.
That's only half the reason it's expensive.
The other reason is that it would likely take years to spend $4000 (plus the real cost of electricity) worth of tokens on a 3rd-party provider that's running a similar limited, DS Flash type model. By that time, the hardware will be obsolete, assuming it's still operational.
> it would likely take years to spend $4000 (plus the real cost of electricity)
Since that cluster only yields 20-30 tok/s on that size of model, at least a decade before the hardware breaks-even with current token costs, and that's not counting electricity. Assuming continued downward pressure on token prices, and the cost of electricity, it never pays for itself.
As a counterpoint, my homelab/home-LLM hardware has appreciated in value by about 60% since I bought it.
Of course, it's not real unless I sell, and the value will eventually go down, but so far I have significant paper profits.
Also, DeepSeek token prices are continuing to _increase_, not decrease.
> DeepSeek token prices are continuing to _increase_
One increase does not a trend make. And the current crop of models are now undercutting deepseek flash...
It's "insane" compared to the $1,500 it should have cost before the RAM crisis
I am pretty confident that given a $200 subscription on any of the big labs, you're getting $4000-$8000 per month in subsidized tokens... do what you wan't with your dough... and I too have a spark that I got really early (October 2025), but no, economically it does not compare to what's runnable locally in terms of quality from the frontier models. Economically, it looks like for as long as there are subscriber plans, you're better off renting.
Before getting the spark, I was just using a google colab account, their $49 dollar plan allows you access to h100's and I can run qwen there in a Jupyter notebook... and if I really need that web front end I can just use cloudeflair/tailscale/the local ssh client to reverse tunnel it.
The cloud stuff is definitely a much better economic value, but I would argue:
1. You learn a lot more running this stuff yourself (especially since you can poke at its internals if you're interested or watch the reasoning chain.) Just being a consumer of this stuff doesn't really teach you much about it other than model & harness specific tricks that become obsolete pretty quickly. (IE, your Claude.md from 6 months ago probably needs a rewrite). Which is fine, I don't think you're going to be "left behind" if you're not a hardcore AI enthusiast or anything (I'm not), but as a guy that's always been interested in computer science I want to see how it ticks.
2. You can't really depend on this subsidization lasting forever IMO. I know the financials thing has been beaten to death but I guess I'm in the camp that it's good to be in control of your tools so that you can go elsewhere if the economics change.
I like to check in with ccusage pretty frequently, and honestly like if I were paying API prices for Claude I'd probably be paying thousands a month.
3.privacy
Any organisation or individuals not wanting to have their sensitive data flowing away (either because of trade secret or data protection laws)
I am also not sure I would choose to use the cheap and easy to run at home model, given a choice. The marketing copy says this is a frontier model, but it's not. Sol and Mythos are the frontier right now. GLM 5.3 Flash simply isn't. I'd rather use the frontier model as they waste less of my time than even Opus.
I thought 4000 in sum. No wait, 4000 per, plus tax. Or EUR pricing to similar accord. Ouch.
Yeah, that little cluster costs about the same as a brand-new Dacia Sandero.
to save others from having to look up what that is like I did, it's a car
Yeah, yeah. BUT, will the Sandero be ... load-bearing? :)
It can bear the load of a few people, at least.
Yes, but it was $200 off!
~$4000 USD each on Amazon, $175 for the cable.
The cables are ~USD $50 from AliExpress although I'm not sure what the tariff situation is for Americans (I think I paid $75 all-in CAD for them)
I mean, I have the same machine and the pricing is only what it is because it has that 1TB nVME in it instead of larger. nVME prices are insane and have been for months.
Reality is on a single spark I'm constantly running out of room and it being an odd size M.2 slot it's a pain to upgrade. I'm setting up a NAS over RDMA via ConnectX though, that's fun.
I already have a recycled QNAP-now-TrueNAS that has 10G connections into the fleet so the 1TB doesn't bother me at all. I did some rough math and I don't think I'll ever need to load weights off NFS for what I'm doing so far, but the capability is there.
> I would love to see an Opus-4.8-level local model but TBH I just haven't got there yet. The models I've tried so far _are_ good but they aren't able to solve tough technical challenges, regardless of harness/prompting/etc.
Agree. It doesn’t even have to be local, using models in this size class through OpenRouter will reveal their limits if you work side by side with Opus level models regularly.
There are a lot of social media posts about people cancelling their Anthropic or ChatGPT subscriptions after installing a local LLM. I’ve used local LLMs a lot and I spend a lot of time with frontier models and the difference is still huge. As far as I can tell, the social media posts about local LLMs replacing frontier models are either wishful thinking, engagement bait, or people who must be working on much simpler projects with a much higher tolerance for slop than I have.
I have exactly the same opinion
Over the last couple years I’ve had to learn sales and understand the thought process behind this better, and I think I’m beginning to understand it
The psychology is that most people aren’t really trying to optimize for productivity (even most people who think they are) on an ROI basis, because their compensation is too decoupled from their actual raw output, and more closely coupled to how differentiated their marginal contribution is to peers. They’re much more incentivized to spend their personal/work time optimizing for being more skilled or acquiring some kind of competitive advantage relative to baseline.
Most people don’t consciously run the numbers of “I get paid $X/hr to add $Y of value” or model pay at work as something with variable inputs (eg something that can be increased with high performance), so it makes sense to them to spend 20 hours of time to save $100 or to make themselves 5% less efficient to take home 0.5% more or avoid doing something they don’t want to start doing.
NOT saying this always happens or that they’re stupid for doing so. I didn’t even realize how much I had been doing it myself until I started recognizing it, and shifted to having my own comp/performance fully aligned with the company’s P/L.
It actually makes a lot of sense IF you can accurately estimate incremental upside (which is much harder and more diffuse than modeling downside if you’re salaried a employee) or if the upfront skill/knowledge investment that looks like bikeshedding pays off in the long run.
These are great points. It's a little off topic but what you bring up is why i advise new grads to spend the first couple years of their career in small eat-what-you-kill companies. I think software devs who start out in large companies get this distorted view that their twice a month direct deposit is just magic and comes from the ether no matter what they do. The whole industry would be better off if everyone started out in a "you don't deliver, you don't eat" company and grew from there.
To be fair, there is no 3 turns that I don't have to jump in into what Opus 5 is doing. There is either some regression or my prompting skills are so much worse now. Flash is not perfect and honestly some things depend on how big context do you keep. So I'm keeping like a really short context with my flash, but it works okay, even though it has a tendency to overthink, and yeah, I run it always in max effort mode.
If you used the bare API pricing, 1M tokens @ 30% input/70% output/50% cached, you'd pay $0.05805. Even with four discounted sparks, how much are you paying for the same tokens/distribution?
There's soooo much by way of experiments, explorations, tinkering, and even projects that you can't possibly pursue through a some SaaS API.
The more reasonable comparison is against rented GPU's, while looking at tradeoffs in latency and upload/download/storage/instance management overhead.
Buying hardware for local models is meeting a wholly different need than buying tokens through OpenRouter or whatever.
For me it's entirely because I have a bunch of projects with my own personal data that would be tough to do with openrouter/claude or any other cloud.
For example, I have a small posix-shell-based LLM harness that can SSH into my NAS and run organization tasks using the local DS4Flash that I have right now. It's already been a massive help for me to keep me organized, and that's just 2x DGX Spark's worth of compute.
It cuts both ways. A GPU in your basement is a depreciating asset with fixed computing power and consumes electricity. Switching model providers is trivial.
> A GPU in your basement is a depreciating asset
All decades prior and up to about a year ago, I would have agreed with you. My Framework Desktop, however has appreciated in value by 75% since I bought it. Will it stay there for a long time? Probably not. But it shows that there are no hard and fast rules about things anymore.
Chinese labs are so used to manipulating benchmarks to try to flatter inferior models that when they finally have one that's really pretty good I think the official announcement here undersells it.
https://deepswe.datacurve.ai/
That's pretty solid. Smarter and cheaper than Luna xhigh, not as smart but less expensive than Luna max. Smashes deepseek v4 flash, and even worse it matches v4 pro at a tiny fraction the cost. Roughly equivalent to sol medium, at a fraction the cost.
They should've just lead with real, up to date data, because it's good, not the silly old tactics like comparing to Opus 4.8 when 5.0 is out in many of their charts.
Congrats to them!
Maybe others have found otherwise, but I find the benchmarks drastically different to real world "feel" of a model, even within the same harness. I'm not sure if this just reflects personal interaction styles, or if it is indicative of benchmaxxing or unrealistic automated benchmarking methodology.
Opus 5 consistently comes at or near the top, but outputs constant unreadable jibberish. Meanwhile GPT 5.6 Luna medium tends to be rated pretty poor on agentic tasks compared to the Chinese lab open models, but I find the latter much more likely to lose track of their own behaviour during a long-horizon task or get stuck in a doom loop.
(This isn't a comment on GLM-5.3 Flash as I've not used it!)
I've been using 5.3 since they initially announced it and my gut feel is that it's not as good as 5.2 for agentic tasks. I'm still using it - I don't think it's bad. I'm just not convinced it's better.
It's also better than Sol (at whatever effort) at designing pretty UIs. I have a Codex sub and I've been using this model for UI stuff.
Only 73K output tokens too. Anthropic should really be embarrassed with their Sonnet 5 price/performance.
Opus 5 is better than Fable in this benchmark?
Even Artificial Analysis has Opus 5 better than Fable in their aggregated "Intelligence Index" which combines 9 benchmarks. Opus 5 is heavily benchmaxxed.
> They should've just lead with real, up to date data, because it's good, not the silly old tactics like comparing to Opus 4.8 when 5.0 is out in many of their charts
It's what people know. Opus is just the common target.
> Smarter and cheaper than Luna xhigh, not as smart but less expensive than Luna max. Smashes deepseek v4 flash
The problem with this and DeepSWE is it goes for a very specific profile. I'm not convinced DeepSWE is any accurate in actual work. It's surely a different signal (compared to some that allow cheating) but it has its own issues, e.g. weak harness.
Luna is great at following instructions but bad instructions or anything not covered = death.
Deepseek is more analytical. Good for bug tracking.
GLM is a better all rounder in some ways. Better at creativity.
I was quite surprised that Zai had deep pockets to serve this free for a week. My first guess was this was an American lab like xai or google
You guys read Z.ai's terms of service, right?
Broad and perpetual license over inputs and outputs, and even your name and profile picture.
Vague prohibitions on whatever may harm Z.ai’s "interests" or even the "national interests" of any country.
Vague prohibitions on "disturbing" or "inappropriate" content, whatever that is.
Vague prohibitions on discussing Z.ai, even my posting this comment violates it.
Can ban you if you, in the "sole and absolute opinion" of Z.ai, have violated these broad terms, and if you paid for the discounted yearly plan kiss your money goodbye.
Isn't this practically every TOS though?
Nearly every TOS I've ever read has a "We can ban you for any reason, or no reason, are under no obligation to disclose any reason." line somewhere in it.
HN's for example
> We reserve the right, at our sole discretion, to change or modify portions of these Terms of Use at any time.
> You acknowledge that Y Combinator may establish general practices and limits concerning use of the Site,
> You further acknowledge that Y Combinator reserves the right to change these general practices and limits at any time, in its sole discretion, with or without notice.
> Y Combinator reserves the right to investigate and take appropriate legal action against anyone who, in Y Combinator’s sole discretion, violates this provision, including without limitation, removing the offending content from the Site, suspending or terminating the account of such violators and reporting you to the law enforcement authorities.
[flagged]
This is not uncommon.
Also, it only applies to their chat offering, not the api. OpenRouter also offers the API with ZDR.
While shitty, i’d say that its really not that special.
???
None of the major LLM chat providers (ChatGPT, Claude and Gemini, and I just confirmed this) claim rights over your input.
They also don't claim rights over your output, but because of how copyright law might apply, they explicitly assign all the rights to the generated output.
Not just that but, even if they wanted to claim ownership of the output, courts in the US have deemed that copyright cannot be assigned to machine-generated output.
So yeah, happy to take a look at a counter-example if you have one (aside from GLM 5.3, obv.).
> In choosing to submit, create, generate, record, post, or display Inputs on or through the Service, you grant an irrevocable, perpetual, transferable, sublicensable, royalty-free, and worldwide right to SpaceXAI to use, copy, store, modify, process, adapt, transmit, distribute, reproduce, publish, upload, download, display in public forums, list information regarding, make derivative works of, and distribute such Content, including anything referenced therein, in any and all media or distribution methods now known or later developed, for any purpose, and to aggregate your User Content and derivative works thereof for any purpose, including but not limited to: (i) maintain and provide the Service; (ii) improve our products and the Service and for our other business purposes, such as data analysis, customer and market research, developing new products or features, or identifying or displaying usage or User Content trends; and (iii) perform such other actions to enforce these Terms, comply with our Privacy Policy, comply with applicable law or governmental, court, and law enforcement requests or requirements or keep our Service safe.
> To the extent the User Content includes a person’s image, likeness, voice, or other similar attributes, you grant SpaceXAI the same rights to use those attributes as part of the User Content as described above. You represent and warrant that you have obtained all rights, licenses, notices, permissions, and consents necessary for SpaceXAI to use that User Content.
https://x.ai/legal/terms-of-service
These are arguably even worse to be honest. Absolute nightmare.
I get all that.
Then alternatives are:
- Grok - where I absolutely have 0 trust in X.ai's interst in "pushing humanity forward".
- OpenAI and Anthropic - which seem to try to be building the biggest moat they can by pushing to ban open models. And at the same time want to be an Arbiter of what level of intelligence I can use.
- Google and Meta - I don't need to talk about the practices of these companies.
Yes, the terms of service aren't great. But the alternatives aren't great either. I don't believe that a future which OpenAI and Anthropic are pushing for has my best interest in mind.
Or Deepseek, Qwen, any other open model hosted by whoever you trust most.
All the American companies you mentioned still follow American law and regulation. Skirting that blatantly has big consequences.
Chinese companies do not follow American laws and there are absolutely no consequences for violating it.
Moreover, the average American is not even aware of exactly what the legal/judicial environment is like in China. If your code and data is stolen, you can't fly to China and demand justice in the courts.
The model is MIT-licensed, so run it on any of the non-Chinese inference providers that will host it in a few days.
It's China. It's a given that they use your data for training. At least they're nice enough to be honest about it.
It's not like US companies don't do the same either.
The model weights are MIT licensed.
Give it a couple days, and there will be plenty of other inference companies hosting it. Don't like z.ai's TOS? Use the model on a provider with TOS that you agree with.
I blocked Z.ai as soon as they were loading 10 different external providers including Alibaba who was just proven to execute silent sound fingerprinting mechanisms.
None of that applies if you run it at home. Also 3rd party providers will start serving this pretty soon under different terms.
They all do that, some are just more honest to tell you upfront than others.
TOS is and will ever be just a "pretty please".
Who the hell cares when I can run it myself?
Chinese laws are not valid in the EU
That’s pretty funny to say when the EU claims GDPR applies worldwide.
This is going so fast! What a time to be on hackernews:
July 16th: The "Kimi K3 moment" - China has caught up to Opus!
4 weeks later: GLM 5.3 - Same performance, but cut the amount of parameters and cost to a third!
12 days later: GLM 5.3 Flash - Almost GLM5.3 performance but cut the parameters in half, cut prices to a fifth and serving on Chinese chips!
With tiny models surpassing huge, 6 months old models on benchmarks, does anybody have some smart words to share on how these still "feel" different?
Artificial Analysis ranks GPT 5.6 Luna similar to GPT 5.4, but that never matches my real world experience. AA seems to do a good job making a single number as representative as possible but there is still so much benchmarks don't communicate.
I agree. They are definitely good - no issues with instruction following for example - but they miss the "intelligence" larger models have.
For implementation tasks, where I have the problem already defined and researched, or just simple task, I'd definitely use something like Luna xhigh or max. If the task is vague, or involves planning, I'd rather use Sol medium, even though it's theoretically worse on benchmarks.
Will we need all the data centers being built or will improvements in software and hardware allow the majority of AI workloads to run locally or in the cloud but way more efficiently than was projected when all the plans were laid out?
Like were executive at Google and AWS and Microsoft expecting this kind of performance from models smaller than what openai/anthropic have been doing? Are we really in a "compute desert"?
> with all of this traffic served on Chinese AI chips
RIP Nivida shareholders
I don't see a situation where subscription payers move outside American LLMs (chatgpt, claude, gemini)
And I don't see a situation where serious API payers are OK with handing the Chinese state all their data. Like manufactures of decades past did and learned a hard, even existential, lesson for it. The state mantra has been "Collect and Copy" for a long time now, tech just hasn't had that moment to experience it yet.
So that leaves local hosting/leasing, but one of those has totally non-practical economics and the other doesn't have enough compute to meet any kind of real demand.
I also have yet to meet a single person who isn't neck-deep in the tech space mention a Chinese LLM. It's 100% the big American three.
If anything it's custom chips from the labs that threatens Nvidia.
What about the current situation, where serious API payers are increasingly OK with using open-weight models running on US providers?
https://www.ft.com/content/32a70a3c-7d28-40b4-808e-36edb58c7...
I can easily see a situation where most non American AI usage is on Chinese models on Chinese chips though.
Genuine question but who do you put as the "three" in big three.
Because I genuinely can't tell if you mean Google or SpaceX/X.ai lol.
Google probably serves more tokens then OAI and Anthropic combined, even if many of those tokens aren't from explicit gemini requests, but from AI overviews and other service integrations.
xAI is already selling spare compute, and basically exists just to gas spacex's perceived valuation.
This is the takeaway here: That's how they have been serving it at scale as Ox-Alpha. This is a definitional moment.-
Further quote:
"Compared with our initial baseline on the same hardware, we achieved a 3× improvement in end-to-end serving performance, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs. This demonstrates that Chinese chips can support frontier-model inference efficiently and economically at scale."
https://z.ai/blog/glm-5.3-flash
Another self-inflicted own courtesy of US government policy.
While I think China would always get to hardware self-sufficiency eventually, all export controls have done is (1) accelerate China's development, and (2) divert revenue that would've otherwise gone to NVIDIA/AMD/etc instead.
Long term it's irrelevant. The only relevant thing is that there's lots of money in chips that can do high performance inference. You see all kinds of competitor products in development or already on the market even here in the US where there are no such restrictions. Cerebras comes to mind. It's natural and expected that eventually Nvidia will either have to keep way ahead or competition will catch up with specialized products.
That doesn't mean by any stretch of the imagination Nvidia will disappear. But the entire stock market valuation, not just tech, has had me scratching my head for a while.
The export controls were revoked before it triggered Chinese protectionism: https://www.silicon.co.uk/e-innovation/artificial-intelligen... / https://archive.vn/B2pah
Revoked or not, just ever having those controls signals to the Chinese ecosystem that you're not necessarily a reliable supplier (Would you trust US export policy to remain stable for the next ~decade given the state of US politic?) and to the Chinese government just how strategically important you see these components.
This isn't the kind of thing you can hash out in public and go back and forth on. Once you put it out there, the other party will take steps to make sure they don't have to rely on us in the long run.
The export controls were not revoked, only reduced, and not before, but after China refused to buy low performing chips. Top gear was and is still sanctioned, as is any EUVL equipment.
And to add to the above: by building their own supply chain for chips, China is helping the unprivileged, those who can't front-run the market with long-term contracts. If China wasn't producing their own chips, the prices for us would be even higher.
Similar to the war-pricing of oil, China's reduction of imports is actually helping to keep our inflation from going even higher.
> The export controls were revoked before
Zai is on another "export control" list outside the broader 1. Doesn't help.
Ox Alpha is a smaller model and it was running very slowly. Chinese AI accelerators are coming along, but nVidia’s lead is huge.
It was being served for free. They were almost certainly being overloaded.
Has there been any confirmation about what that model even is?
Edit: Ah:
> This stealth model was developed and operated by ZAI, revealed to be ZAI GLM-5.3-Flash.
It's also in this very announcement, in the first paragraph:
> Before release, we tested GLM-5.3-Flash anonymously as ox-alpha on OpenCode and OpenRouter to gather user feedback. It quickly became the most popular model of the week — with all of this traffic served on Chinese AI chips.
Not really a brag: it ran like shit. Very slow (~20tps, VERY high latency) and it would timeout all the time.
I'm sure the chips are fine, but they clearly didn't have enough capacity for the demand they had (that 100T/day claim was asbolute bs)
seems unlikely that they'll get nearly as much demand now that it isnt free
Sure, although I still expect it to become the most used model on openrouter.
God I wish I could’ve shorted NVIDIA right now
Whats stopping you? You could buy puts right now.
Get a 210 strike put contract and if your thesis is that nvidias current 10 day slide continues you could make some money.
Unless NVidia craters you are likely to lose money given the IV crush that will happen today.
It's earnings day for them...
Which 9/10 times hasn't been great anyway (stock reaction).
Of course the release was not coincidental - with the earnings days - I am sure.-
And two models released same day + openai chip
yay, I called it! :) (in the other thread)
This is no surprise [0] [1].
>> "They are already there on open weight models and Jensen knows that it is only a matter of time until China catches up with GPUs or other AI accelerators."
It is also why Nvidia becoming a bank for other AI companies who are unable to find VCs to fund them isn't really a good thing and that is bearish.
[0] https://news.ycombinator.com/item?id=49397204
[1] https://news.ycombinator.com/item?id=49431231
Is the actual Z.AI ecosystem good enough to replace the main drivers like Codex and Claude? Because it looks like Z Code is just a Codex fork. Just like the Kimi Code one is.
What irks me about this is that the harnesses seem to be just an afterthought here.
Don't get me wrong, I love messing around with installing Pi, getting it hooked up with OpenRouter, and just trying all kinds of different stuff, local models, etc... but when it comes to literally just setting up a productivity environment and trusting my entire machine with it, I just run Codex.
I have heard from anecdotes where people have indeed replaced their main drivers with DeepSek V4 Flash or GLM and state that "it's almost as good as... [claude/gpt]" but I never hear anyone say "yeah, this is the model/harness that I now run on my machine and don't mess with it"
> "yeah, this is the model/harness that I now run on my machine and don't mess with it"
* me raises hand.-
Their list of allowed tools is extensive so just use whatever you want within that list
Think z code gives a token bonus though
> 320B total parameters and just 18B active parameters
This is pretty hefty for a "flash" model, even a 256 GB setup is insufficient at q4 - and q4 is already the worst-but-still-acceptable quant in my experience. The benchmarks look great, especially since GLM tends to be more honest than the average Chinese lab, but you’ll need to splurge to run it at home.
@edit: so many releases that I forgot to math. This fits just fine in q4, realistically the minimal hardware would be 192gb - so blazing fast on double rtx 6000 pro and usable on 256gb unified memory. You could even go with 5bit quant on 256gb.
… you’ll still need to splurge, though.
That's 160GB-ish for Q4...how is 256 insufficient?
Looks like the M5 Ultra Studio wait times are going to increase again. Already at 10-12 weeks, I wonder how long it'll go?
I guess like the M3 Ultra, at some point normal customers won’t be able to buy it.
> it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks.
From a biased source, but would be big if true. I've had great results with GLM 5.2.
From their subscription page, the smallest plan gives you about 97M tokens weekly for 5.3 but 292M for 5.3 Flash. Not exactly 10x the limit.
The recent and slightly smaller DSv4 Flash is also GLM 5.2 equivalent (or close enough)
DSv4 hallucinates much more than GLM-5.2 though.
> From a biased source, but would be big if true. I've had great results with GLM 5.2.
It's at least close (even if not better) from the Ox Alpha runs. For the price it's definitely great.
For those who didn't read, this is the identity of the mysterious "Ox Alpha" model
They even give this over the API now:
│ https://openrouter.ai/api/v1/chat/completions model: stealth/ox-alpha auth: OPENROUTER_API_KEY status: 404 Not Found response: {"error":{"message":"Thank you for participating in the Stealth Ox Alpha testing period. This model was ZAI's GLM-5.3 Flash.
│ Use it now: https://openrouter.ai/z-ai/glm-5.3-flash","code":404},"user_...":"}
Yeah, made me suspicious of how well the Ox Alpha was performing that it wasn't some 'new group' making the model.
When reading this type of announcements, always have keen eyes on graphs.
e.g. "Agent Coding Performance by Effort Level" cuts Y-axis from 0~20.
- This makes it as if GLM-5.3-Flash made a bigger jump than it claimed as the Y-axis does not increase much (stupid trick used in biz reports)
I did mention that ox was working ok for me, and having an open-weight comparable to close to SOTA makes it very compelling for me to try it out locally (well, only if I got more VRAM)
they also conspicuously omitted GPT 5.6 Luna from comparison. It scores lower, but is also cheaper. MiMo 2.5 is not a valid comp at this point
edit: nevermind. it is there in the artifical analysis scatter plot, but is greyed-out.
MUCH more interesting is that in that chart, their cost is WAY off. The actual chart shows GLM 5.3 Flash at $0.09, but their chart shows $0.045...
The web page says 5.3 flash is discounted right now.
Seems disingenuous to draw frontier graphs with starter pricing.
Well, Luna debuted with 5x higher pricing than is currently available. With the pace of recent development these models might not be relevant by Thanksgiving.
Of course. Pricing is always changing, but typically it goes down over time, not up. So, if you're showing artificially low pricing from the start based on a teaser rate, IMO, you shouldn't be using that to show where you appear on a frontier graph. Place yourself on the graph based on your expected long-term pricing. Then, over time, adjust your position based on your standard rate, whatever that might be. Games are always being played for things like this, but this seems excessive.
I don't know if that's the standard pricing for US models to go down overtime, while Chinese ones go up (start cheap but pay more).
I don't have enough metrics to compare those costs but still Chinese models have been cheaper except against Luna for me.
FWIW, Luna does everything so well, I just keep using it for all my agents by default.
The key difference between this and all other GLM models is it's multimodal. You cannot send images to the other GLM models.
I really wish GLM models had vision capabilities. I've worked around that in the past to use a vision MCP in my harness that GLM can call. It is not the same, but it allows the model to query images.
Well, now one of them does!
At the current 50%-off GLM-5.3-Flash price ($0.075/M input, $0.25/M output; cached input $0.015/M), surprisingly, roughly $400–900/month would buy token throughput comparable to fully exhausting Claude Max 20×
On OpenRouter the pricing is: Input $0,075/M - Output $0,25/M - Cache Read $0,015 /M
How is the business model of Anthropic/OpenAI will sustain?
They're obviously in a pickle, nobody is going to continue to pay $15-50 a mm tokens here soon. There's a reason OpenAI stopped training large models last week, and it's not because of "saftey" or "alignment" they know these gigantic models are not worth the squeeze.
If we fast forward say 5 years, I don't see how we don't end up in world where people (and enterprises) are more savvy with how they use LLMs. Meaning, more models, smaller models, weirder models, more specialized models, etc. And all of it running on a variety of hardware (edge devices, personal computers, on-demand cloud compute).
I don't see how NVIDIA can keep their spot as belle of the ball. If LLMs and friends are truly to become as useful and ubiquitous as everyone thinks they will, then commoditization is the only option.
We need to figure out what the real pricing is for a going concern. Right now, everyone is subsidizing and discounting to grow (or maintain) market share. The big question is whether the steady state, market derived inference pricing is above or below what we’re seeing today. I honestly don’t know. Anthropic had said that inference is profitable, but they’re clearly not yet profitable overall with training and buildouts still happening.
In my brief testing, it did about as well as Qwen3.8-4B-Distill, and LFM2.5-2.6B overtook both.
It's only 320B, local frontier AI is getting closer, sooner than expected.
Can't come soon enough!
I'm starting to think that this whole sanctioning China may motivate and prompt them to do more and better in every field.
It's too big, bright and resourceful of a country to choose confrontation instead of collaboration.
Starting? This was obvious way back in 2019, when the US decided to give China a little push developing their own silicon industry.
> I'm starting to think
That's good. Keep going.
This has been clearly stated as what would happen going back several decades at least.
Well the big problem with china is that they do not respect international law when it comes to technology theft. But that argument is very weak when it appears that a lot of what they do is out in the open for anyone to replicate.
That's how you catch up when you're behind.
Now the US is behind in EVs can you guess what they're doing? [1]
[1] https://evwire.com/p/video-ford-ceo-jim-farley-says-they-fly...
"argument is very weak" regardless as I said.
There isn't one global "international law" for copyright. There are treaties that countries negotiate with each other.
If the USA wanted a copyright treaty with China bad enough, we would negotiate one. China is not breaking any laws here, international or otherwise.
No major power respects nor cares about international law.
Intellectual property is part of WTO agreements but enforcement is domestic.
US companies do it too, regularly, they simply hire and poach staff from competitors.
Proving it to be IP theft is difficult unless you can prove documents being passed. But often all you need is the know-how of the hired talent.
yeah, America is totally out there respecting international law.
"problem" indeed.
Related: https://news.ycombinator.com/item?id=49446422
(281 points, 118 comments)
Artificial Analysis benchmark is out: https://news.ycombinator.com/item?id=49450353
offtopic: Is there any chance we could see competing models from other countries in the next 5 years?
> Combined with our latest 30T-token multimodal pre-training corpus [...]
Is the optimal formula still 20x the amount of model params in tokens for training? Could this mean we're getting a GLM with 1.5t params?
This was the ox-alpha model, right? I remember it performed really well for a model that had 'flash' in its name.
so is it confirmed if this is the mysterious OxAlpha model?
Yes; if you try to use Ox Alpha it will give an error saying it waa trial period, and that it is GLM 5.3 flash.
from the article, pareto frontier for open source models is completely dominated by GLM now.
Well, it will be interesting to see where Qwen3.8-Flash-Next ends up landing, also released today. These are exciting times!
I find GLM's idea of fast/flash is not really competitive with the speed DS4 Flash has, and it's hard to see them as being in the same segment for that reason.
Standard API Pricing for GLM-5.3-Flash (per 1M tokens)
- Input: $0.15 - Output: $0.50 - Cached input: $0.03
Is that cheaper than DS4 flash?
Slightly more expensive than the (post-price hike) DS4 flash pricing, but in the ballpark.
https://openrouter.ai/compare/deepseek/deepseek-v4-flash-073...
Hm. GLM is more expensive in all dimensions than DS but it has a lower weighted average input? How is that?? Something seems off.
EDIT: Looks like they are swizzling around the pricing dynamically on that page, on both the GLM and the DS sides, so who knows.
Comparison should be to 0731
updated thanks
All I can say is that even if it is, I was almost glad to go back to using DS4 Flash. Because 0XAlpha was just so friggin slow to complete a task because of the level of circular reasoning that it would go over and over into, sometimes even returning no output. If I just wanted something done I would switch from a free model to a paid one which is crazy.
Tbh it was also slow because it was being hammered by everyone making use of the free tokens
Possibly influenced by that, but I believe that is a different issue. I meant the way it processed a request. It went into so many more loops of thinking.
It's even cheaper than DS4's off-peak pricing. Seems like DeepSeek have some stiff competition now
Few weeks ago, I wouldn't expect this statement to be true. Accelerate!
How much is the “discounted” pricing they mention?
Why is their own coding plan always the last place z.ai release their models? Its even online, you just have to guess the model settings.
Any providers hosting it outside of China?
Anthropic is accelerating their IPO cause they know what's coming in the next 5 years.
> To overcome the relatively limited compute and memory capacity of individual chips, we built a dedicated inference engine for this architecture on top of SGLang. Notably, this effort was accelerated by our GLM-5.3-powered infrastructure agent, which assisted engineers in developing and optimizing kernels, diagnosing performance bottlenecks, and improving the serving stack — creating a feedback loop in which the model helped optimize the system serving the model itself.
> (...) Compared with our initial baseline on the same hardware, we achieved a 3× improvement in end-to-end serving performance, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs. This demonstrates that Chinese chips can support frontier-model inference efficiently and economically at scale.
It might be one of the most actually practical tasks that AI might've done because the compounding effects of it and also its implications are/feels so immense. It feels as if Nvidia might be in a slight turbulence from it.
This looks like it goes hard, can't wait to try it
By clicking this link you download some PDF in the background
Its displayed in the html...
I didn't accept a single edit from this model over the entire week, just saying. I do not understand how it's being benchmarked on par with Sol and other larger models.
[dead]
[dead]