I maintain a fork of ds4 as shared libraries and thus can be used with other languages via FFI, along with public builds/binaries [1]. I made ds4go [2] against ds4 using techniques inspired by yzma.
In addition to the library bindings, we have a small library of tools (workspace for view/edit, scratchpad for persistence) and making your own is registering a Go function. And in recent weeks, I added the Vision and Qwen support, as ds4 added them.
Even if you don't use the Go library, the ds4go binary makes it really easy to download the libraries off of HuggingFace with a TUI available vie Homebrew.
Here's some TUI toy screenshots, sorry I still haven't released that code; it's of different quality than the others. [3]
Nothing comparable but inspired from DwarfStar I wrote a little inference engine for Intel Xe-LP (no XMX) 32GB laptops. The only model supported right now is a quantized Gemma-4, but I don't exclude in the future to support other MoE of similar size. Too bad we have no Qwen 3.8 35B-A3B yet.
I'm also looking into expanding the protocol and the engine to support various steering techniques.
Just tried this on my Intel Ultra 7 255H, I also only have an iGPU. This does ~22tps! Love this.
I just had to do a little patch to support my iGPU device that is a bit newer than Intel Xe-LP, maybe I'll do a PR.
On a side note the other day I was experimenting with Sonnet 5.5. I gave it the llama cpp repo and told it to extract in a single file inference for a single model + backend (qwen3.5 4b mtp + sycl) and (after a long time) it actually worked! It produced a ~1400 lines file with no deps. I need to check the quality of inference yet but I think this is still a great achievement.
I'm pretty sure 2027 will be a very interesting year for local models and inference.
Yeah I see the value but I built Xenolith to target smaller models.
I heard antirez saying that he designed DwarfStar also to be forked and tuned to everyone's specific needs. Do you have a specific machine/spec in mind?
Ive been using this since it was initially released with deepseek v4 flash, and it is absolutely the best launcher ever on my m5 max 128gb
Now Ive been running qwen 3.8 flash next for more than a week and it’s doing great, really fast and super long context windows. Sometimes the model is behaving stupidly by not remembering something I said earlier but it could be also a problem from the agentic AI harness. Im using oh my pi but Im wondering what people are using ds4 with here ?
If you are interested in high-end models with high-end Apple Silicon, also try out Local Code: https://releases.drawthings.ai/p/public-beta-of-local-code-b... It is currently in TestFlight (and will open-source next week), supporting vision with DeepSeek 4.1 Flash, Qwen 3.8 27B and DeepSeek 4 Flash 0731 without vision. Custom quants & SSD streaming to make these big models work with 64GiB and above devices (and of course, Qwen works with devices with 16GiB and above).
Metal, the primary target, on Macs with 96 GB or more. Smaller machines can use SSD streaming. SSD streaming is also needed in order to run very large models such as full GLM 5.x (not Flash) on 128GB systems
Anyone tested token speeds at less than 96gb RAM on apple?
small native inference engine optimized first for DeepSeek V4 Flash (including the experimental vision model), DeepSeek V4.1 Flash (Metal, and text inference on CUDA), and additionally GLM 5.2 and 5.3, GLM 5.3 Flash and DeepSeek V4 PRO, and Qwen3.8 Flash Next (Metal and CUDA)
This is local targeting high end consumer hardware like DGX Spark or AMD Ryzen AI Halo.
For our mere mortals that were kids not long ago and can't really believe we've got our hands on a x090 series targeting Qwen3.8 27b, https://github.com/noonghunna/club-3090 is the way to go.
First of all, a x090 series card is "high end hardware" in its own right these days. Secondly, I'd think you'd probably get more interesting results running a MoE model in CPU-MoE mode, i.e. with the shared parameters residing on GPU and sparse experts on CPU plus SSD offload. Yes it will be slower, but small dense models are just a dead end and not that interesting. (Note that prefill would still be sped up in this setting; the CPU/GPU layer split in llama.cpp and the like applies to decode, but even a "0 graphic layers" setup does accelerate prefill.)
Antirez understood the fundamentals of how LLMs work, and read the inference code (or summaries of it via ai).
But what really made the difference was his understanding of hardware and systems programming in general and low level or architectural tricks to pull.
It is pretty nifty. I spend some time over last weekend implementing fused TQ to allow for 1m context lengths on a 128 gb MacBook M5 Max when using Qwen 3.8 flash next (https://github.com/antirez/ds4/pull/1115 if you are interested). If I get bored I might port over the Metal kernels from oMLX -- the speed increase they have for the v0.7.0 release is amazeballs.
It's more likely to work. Most LLM runners are meant to work with any model, which means there are all kinds of ways you might misconfigure them in a way that causes function tooling not to work, or performance to be less than you would like.
DwarfStar's selling point is that it only supports a small set of carefully chosen models, but it supports them really well.
I'm sorry, it's hard for me to understand what point you were trying to make. So existing LLM runners are designed to support all models, and they run all models, but they might be misconfigured? And DS4 is better because it's unable to run all models?
Emphasis on performance and usable coding/agentic ability for consumer AI hardware. Does not attempt to handle all models or hardware at once but rather focuses on optimizing the best options for that category of hardware.
Lots of small details are taken care of so it runs smoothly. For example ds4-agent is append only, never rewriting history of messages, keeping KV cache prefix reusable. Huge benefit
The creator of this has a very interesting YouTube channel he posts to almost daily, talking mostly about current developments on AI from a technical but also societal/philosophical point of view. I specifically like that he provides a (much needed in this space) leftist point of view while not being anti-AI. Most of the videos are in Italian, so if you speak Italian (or are fine with YouTube automatic translation) I highly recommend it:
There are insane speed improvements for local inference going around on X right now. They've popped up the last month and week.
Tensorfold is getting 100%+ speed increases on both prefill and decode for models like Qwen 27B. oMLX has followed them and have had similar improvements in the past week.
There's lots of different techniques like letting CPU help with prefill, DFlash specualtive decoding etc.
I'm really excited for this as I'll be receiving an M5U in about a month. Expect to be running Qwen 4 27B or Flash (it's a 96gb machine), and they may come close in performance to DS 4/4.1 Flash, and should be able to hit 100 tps. Local is really becoming viable, especially considering that GPT 6.1 has been running at 20ish tps the past week.
Antirez has been writing C for a million years so is much more familiar with it than Rust.
Also the goal of the project is to squeeze the absolute maximum performance and capability possible out of limited hardware resources (compared to clusters of B200s or something).
Does Rust even give you good access to low-level code on different platforms? And if so, how much extra work do you need to do to make it acceptable to the compiler? And is that work worthwhile if you are not going to get the security guarantees of normal Rust code? Is it a worthwhile tradeoff when the goal is performance?
Those are real questions by the way, not rhetorical. If Rust could work well for this type of project then I would like to know.
> Antirez has been writing C for a million years so is much more familiar with it than Rust.
This is explicitly an AI-coded project, Antirez argues that LLMs are worse at writing Rust than C because so much high quality systems code (think e.g. sendmail) that ends up in AI training sets is C, not Rust. Another related argument is that the more detailed syntax and compiler feedback found in Rust compared to C are really a negative for LLM workflows.
There's plenty of room to disagree wrt. this of course: without the strong typing checks of Rust around e.g. indirect references, safety and correctness ends up being a global property in typical C programs, and LLMs are terrible wrt. reasoning about global properties. You're better off forcing them to adapt to a different local syntax that does a more complete job of enforcing modularity, since this is comparatively foolproof.
Yes, Rust gives you great access to low-level code on different platforms, including SIMD. It is also alias-free by default, and gives you excellent primitives to write multi-threaded code with compile-time correctness guarantees, which is how projects such as zlib-rs end up significantly faster than their C counterparts.[1]
It's about as good as it can get for this kind of code.
Personal preference of the author, he made at least one video on YouTube on why he dislikes Rust. I think he finds it too cumbersome and not worth it when the software isn't security-critical (not that I agree, just reporting what IIRC his stance is).
He recently said that he finds Rust less ergonomic and that this also affects code written by LLMs, which he thinks excel at writing C partly because of the enormous, high-quality codebase they were trained on. He sees security-critical code as a reason to choose Rust.
And you know it's load bearing each of the load baerings parts that bear some load and load a bear... you fight a bear because it took a load... or something like that...
Right? Compare this with antirez's blog lmao. The very author of an incredible piece of software using the most plain website possible, while a derivative post about the same tool it's a slop fest with useless FX, cringe hackerman style palette and such. It's just too funny.
Here's a sample of the site's headers. Note the heavy reliance on slop marketing-speak (rule of 3, X not Y, etc.).
"Compressed, not lobotomized."
"Dense, resident, yours."
"Local frontier inference, narrow on purpose."
"ds4 hardware fit: local, streamed and distributed."
The whole site says nothing with so many words. It's also got all the hallmarks of a typical vibe-coded web site (small all-caps text, highly sectioned content, silly animations). Why do people do this? It doesn't impress. In a few years, we'll look back on sites like this like we look at geocities sites today.
He had an awesome opportunity to do high concurrency synchronous replication (raft) on top of in-memory databases at a time where ssds were still uncommon, but instead chose to redneck-engineer his own protocol, then double-down that he knows best.
Not that dissimilar to choosing C over rust for familiarity.
aphyr: distributed systems are difficult and break in ways that are difficult to predict, this is known scientific fact and here's a long list of databases I broke because their engineers think the rules don't apply to them.
antirez: no, u!
elktown: both sides are tribals with superiority complexes!
I maintain a fork of ds4 as shared libraries and thus can be used with other languages via FFI, along with public builds/binaries [1]. I made ds4go [2] against ds4 using techniques inspired by yzma.
In addition to the library bindings, we have a small library of tools (workspace for view/edit, scratchpad for persistence) and making your own is registering a Go function. And in recent weeks, I added the Vision and Qwen support, as ds4 added them.
Even if you don't use the Go library, the ds4go binary makes it really easy to download the libraries off of HuggingFace with a TUI available vie Homebrew.
Here's some TUI toy screenshots, sorry I still haven't released that code; it's of different quality than the others. [3]
EDIT: add ds4go TUI screenshot gist [4]
[1] https://github.com/NimbleMarkets/ds4/releases/tag/v0.8.20260...
[2] https://github.com/nimblemarkets/ds4go#install
[3] https://gist.github.com/neomantra/ae47422c8daf7a458212c93992...
[4] https://gist.github.com/neomantra/40180ade13df93290250ce8c6d...
https://github.com/antirez/ds4
The project GitHub page is a much better introduction for the hn crowd.
Nothing comparable but inspired from DwarfStar I wrote a little inference engine for Intel Xe-LP (no XMX) 32GB laptops. The only model supported right now is a quantized Gemma-4, but I don't exclude in the future to support other MoE of similar size. Too bad we have no Qwen 3.8 35B-A3B yet.
I'm also looking into expanding the protocol and the engine to support various steering techniques.
https://github.com/simoneiacomino/xenolith
Just tried this on my Intel Ultra 7 255H, I also only have an iGPU. This does ~22tps! Love this.
I just had to do a little patch to support my iGPU device that is a bit newer than Intel Xe-LP, maybe I'll do a PR.
On a side note the other day I was experimenting with Sonnet 5.5. I gave it the llama cpp repo and told it to extract in a single file inference for a single model + backend (qwen3.5 4b mtp + sycl) and (after a long time) it actually worked! It produced a ~1400 lines file with no deps. I need to check the quality of inference yet but I think this is still a great achievement.
I'm pretty sure 2027 will be a very interesting year for local models and inference.
Please open a PR! I was too conservative with the supported devices.
If your GPU supports XMX we could also explore using it to improve the prefill kernel, but I don't have the hardware to test it myself.
I wish someone would add Intel support to ds4. And also improve AMD support.
Maybe Intel and AMD should help them with that.
AMD sent antirez a Strix Halo back in June for this purpose
Yeah I see the value but I built Xenolith to target smaller models.
I heard antirez saying that he designed DwarfStar also to be forked and tuned to everyone's specific needs. Do you have a specific machine/spec in mind?
The recent Intel GPU/AI cards. Really the same type of models as ds4
Ive been using this since it was initially released with deepseek v4 flash, and it is absolutely the best launcher ever on my m5 max 128gb
Now Ive been running qwen 3.8 flash next for more than a week and it’s doing great, really fast and super long context windows. Sometimes the model is behaving stupidly by not remembering something I said earlier but it could be also a problem from the agentic AI harness. Im using oh my pi but Im wondering what people are using ds4 with here ?
If you are interested in high-end models with high-end Apple Silicon, also try out Local Code: https://releases.drawthings.ai/p/public-beta-of-local-code-b... It is currently in TestFlight (and will open-source next week), supporting vision with DeepSeek 4.1 Flash, Qwen 3.8 27B and DeepSeek 4 Flash 0731 without vision. Custom quants & SSD streaming to make these big models work with 64GiB and above devices (and of course, Qwen works with devices with 16GiB and above).
For our mere mortals that were kids not long ago and can't really believe we've got our hands on a x090 series targeting Qwen3.8 27b, https://github.com/noonghunna/club-3090 is the way to go.
I'm maintaining a web frontend for this, trying to at least. You can follow it here: https://github.com/gchamon/club-3090-server
First of all, a x090 series card is "high end hardware" in its own right these days. Secondly, I'd think you'd probably get more interesting results running a MoE model in CPU-MoE mode, i.e. with the shared parameters residing on GPU and sparse experts on CPU plus SSD offload. Yes it will be slower, but small dense models are just a dead end and not that interesting. (Note that prefill would still be sped up in this setting; the CPU/GPU layer split in llama.cpp and the like applies to decode, but even a "0 graphic layers" setup does accelerate prefill.)
Qwen3.8 35ba3b is decidedly faster but also dumber, at least in my tests.
> Qwen3.8 35ba3b
Does not exist. You're thinking of Qwen3.6 35ba3b
Can someone explain me what expertise (domain knowledge) one needs to be able to write such model specific inference engine?
The other inference engine are also model by model with a huge switch statement deciding which part to load for which model or are they very generic?
Antirez understood the fundamentals of how LLMs work, and read the inference code (or summaries of it via ai).
But what really made the difference was his understanding of hardware and systems programming in general and low level or architectural tricks to pull.
Wondering how well this does with tool calling. Any one has any numbers or videos or anything using this?
From the github repo it seems like you really don't need a big Mac with huge amounts of RAM but SSD is sufficient.
If this is anywhere near 50 TPS, that would be a game changer in the personal LLM space!
It is pretty nifty. I spend some time over last weekend implementing fused TQ to allow for 1m context lengths on a 128 gb MacBook M5 Max when using Qwen 3.8 flash next (https://github.com/antirez/ds4/pull/1115 if you are interested). If I get bored I might port over the Metal kernels from oMLX -- the speed increase they have for the v0.7.0 release is amazeballs.
7800€ for a MacBook Pro M5 Max with 128GB RAM, right now. Insane.
I’m already able to use 1M context windows with the same machine than you and same model. Strange
Yeah, most of what I did was to add fused TQ support to leave more memory free for other nefarious purposes.
Engineers have entered the building. We've come a long way from people debating whether mmap was safe to use.
How is this different from other LLM runners?
It's more likely to work. Most LLM runners are meant to work with any model, which means there are all kinds of ways you might misconfigure them in a way that causes function tooling not to work, or performance to be less than you would like.
DwarfStar's selling point is that it only supports a small set of carefully chosen models, but it supports them really well.
I'm sorry, it's hard for me to understand what point you were trying to make. So existing LLM runners are designed to support all models, and they run all models, but they might be misconfigured? And DS4 is better because it's unable to run all models?
It has better defaults.
Does that imply that with enough effort spent on fine-tuning the configuration, other LLM runners can achieve similar results?
So wouldn't be easier just to provide a repository of better defaults for each model? Like lsp-config for neovim?
Emphasis on performance and usable coding/agentic ability for consumer AI hardware. Does not attempt to handle all models or hardware at once but rather focuses on optimizing the best options for that category of hardware.
Lots of small details are taken care of so it runs smoothly. For example ds4-agent is append only, never rewriting history of messages, keeping KV cache prefix reusable. Huge benefit
My instinctive reaction from the readme is that it isnt. It's apparently a vibe coded knock off of llama.CPP.
The creator of this has a very interesting YouTube channel he posts to almost daily, talking mostly about current developments on AI from a technical but also societal/philosophical point of view. I specifically like that he provides a (much needed in this space) leftist point of view while not being anti-AI. Most of the videos are in Italian, so if you speak Italian (or are fine with YouTube automatic translation) I highly recommend it:
https://youtube.com/@antirez
His blog is good as well, but he doesn't post there as frequently: http://antirez.com
If creator of Redis is usig AI to write serious software, ordinary folks need to rethink their stance.
There are insane speed improvements for local inference going around on X right now. They've popped up the last month and week.
Tensorfold is getting 100%+ speed increases on both prefill and decode for models like Qwen 27B. oMLX has followed them and have had similar improvements in the past week.
There's lots of different techniques like letting CPU help with prefill, DFlash specualtive decoding etc.
I'm really excited for this as I'll be receiving an M5U in about a month. Expect to be running Qwen 4 27B or Flash (it's a 96gb machine), and they may come close in performance to DS 4/4.1 Flash, and should be able to hit 100 tps. Local is really becoming viable, especially considering that GPT 6.1 has been running at 20ish tps the past week.
curious, why did antirez go with C instead of something like Rust ?
Antirez has been writing C for a million years so is much more familiar with it than Rust.
Also the goal of the project is to squeeze the absolute maximum performance and capability possible out of limited hardware resources (compared to clusters of B200s or something).
Does Rust even give you good access to low-level code on different platforms? And if so, how much extra work do you need to do to make it acceptable to the compiler? And is that work worthwhile if you are not going to get the security guarantees of normal Rust code? Is it a worthwhile tradeoff when the goal is performance?
Those are real questions by the way, not rhetorical. If Rust could work well for this type of project then I would like to know.
> Antirez has been writing C for a million years so is much more familiar with it than Rust.
This is explicitly an AI-coded project, Antirez argues that LLMs are worse at writing Rust than C because so much high quality systems code (think e.g. sendmail) that ends up in AI training sets is C, not Rust. Another related argument is that the more detailed syntax and compiler feedback found in Rust compared to C are really a negative for LLM workflows.
There's plenty of room to disagree wrt. this of course: without the strong typing checks of Rust around e.g. indirect references, safety and correctness ends up being a global property in typical C programs, and LLMs are terrible wrt. reasoning about global properties. You're better off forcing them to adapt to a different local syntax that does a more complete job of enforcing modularity, since this is comparatively foolproof.
Yes, it is AI-coded. But definitely not a one shot kind of a deal.
Much easier to work with a language you are most comfortable with right?
Yes, Rust gives you great access to low-level code on different platforms, including SIMD. It is also alias-free by default, and gives you excellent primitives to write multi-threaded code with compile-time correctness guarantees, which is how projects such as zlib-rs end up significantly faster than their C counterparts.[1]
It's about as good as it can get for this kind of code.
[1] https://www.reddit.com/r/rust/comments/1ixt1ei/zlibrs_is_fas...
Personal preference of the author, he made at least one video on YouTube on why he dislikes Rust. I think he finds it too cumbersome and not worth it when the software isn't security-critical (not that I agree, just reporting what IIRC his stance is).
He recently said that he finds Rust less ergonomic and that this also affects code written by LLMs, which he thinks excel at writing C partly because of the enormous, high-quality codebase they were trained on. He sees security-critical code as a reason to choose Rust.
The video is in Italian but has an auto-dubbed English audio track: https://www.youtube.com/watch?v=sOt0WpQG5eU\&t=526s
Because C is the simplest language that a competent programmer learn just in an afternoon pretty much.
I like C's simplicity so much. The only other language that comes close in simplicity and minimalism is go.
In case you like the store kv to disk so you can resume I keep this branch of llama.cpp that includes that same functionality
https://github.com/alainnothere/llama.cpp/commits/disk-cache...
And you know it's load bearing each of the load baerings parts that bear some load and load a bear... you fight a bear because it took a load... or something like that...
Last time I was using llama cpp you could just do:
or and the load equivalents.That was a year or so ago though...
I’m a little confused
ds4 is referring to “dwarfstar” “4” and references DeepSeek V4 most of the time
but its model agnostic-ish
and benchmarks compared to what? what do these large MoE models typically get in tokens per second?
I’m garnering this is just an easier way to load large models per expert on consumer hardware? as opposed to the hackier solutions?
I’m intruiged. Note that the blogpost says 64gb Macs are good minimums while the github says 96gb is a minimum
This website is pure slop. I'd ask @dang to just link the original repo
Right? Compare this with antirez's blog lmao. The very author of an incredible piece of software using the most plain website possible, while a derivative post about the same tool it's a slop fest with useless FX, cringe hackerman style palette and such. It's just too funny.
Here's a sample of the site's headers. Note the heavy reliance on slop marketing-speak (rule of 3, X not Y, etc.).
"Compressed, not lobotomized."
"Dense, resident, yours."
"Local frontier inference, narrow on purpose."
"ds4 hardware fit: local, streamed and distributed."
The whole site says nothing with so many words. It's also got all the hallmarks of a typical vibe-coded web site (small all-caps text, highly sectioned content, silly animations). Why do people do this? It doesn't impress. In a few years, we'll look back on sites like this like we look at geocities sites today.
>like we look at geocities sites today.
We look at geocities with nostalgia, I guess. Ugly as fuck but made with heart when all this thing of the internet was growing.
This slop shit on the other hand... It's cringe right now.
Apparently I'm the only person to whom "from the creator of Redis" is a warning.
Why? Care to explain?
OG vibe coder:
https://aphyr.com/posts/283-jepsen-redis
https://antirez.com/news/55
finally
https://aphyr.com/posts/307-jepsen-redis-redux (see his comments there too)
He had an awesome opportunity to do high concurrency synchronous replication (raft) on top of in-memory databases at a time where ssds were still uncommon, but instead chose to redneck-engineer his own protocol, then double-down that he knows best.
Not that dissimilar to choosing C over rust for familiarity.
>Not that dissimilar to choosing C over rust for familiarity.
It's LLM age. Just port it to Rust if that is so wrong for you. He is doing it for free, no need for arguing about the language he wants to use
That quarrel was so incredibly petty. Oneupmanship, madness of not conforming with the zeitgeist, cherry-picking, whatnot.
Yeah, as usual with devs; pick a tribe then go to insufferable lengths with the newfound and completely unearned superiority complex.
aphyr: distributed systems are difficult and break in ways that are difficult to predict, this is known scientific fact and here's a long list of databases I broke because their engineers think the rules don't apply to them.
antirez: no, u!
elktown: both sides are tribals with superiority complexes!
I meant you.
What does ds4 offer that projects such as llamma.cpp or ollama haven't been offering for a while?
I mean, I've been using local models on vscode right next to frontier models with ollama for a few months. What's new?
the problem is the dsv4 checkpoint so quantized isn't very good
Which ds4 checkpoint for which model exactly did you test? Don't they have multiple different versions and quantization levels?
not my experience
the ds4 quants were very good beating the unsloth quants https://github.com/michaelasper/benchmarks/blob/main/deepsee...
That's quite the statement - unsloth quants are amazing.
Random comment but the name is funny to me, reminds me of Silicon Valley.
What are we going to name the company, how about Dwarfism 2.0? What happened to 1.0 Jared?
I immediately thought of the unique ring from Diablo 2 named Dwarf Star.
Dwarf Star is better than Dirty Socks, or Dynamic Slinky.. definitely not the worst backronym.
Smallulator