The take away is that this model is a smaller model that competes with Haiku, I would hope they come out with a "Sonnet" competing model, then Opus. I have been wondering why Microsoft is kind of "sleeping" on offering models they themselves have made on Copilot, maybe it was part of their deal with OpenAI? Not sure.
Does anyone actually uses these smaller models for coding? If so, how? I usually Opus everything. Is the play to plan/design/architect with a heavier model than delegate structured tasks to these smaller ones? Would appreciate to hear someone's opinion on having done and tested both paths.
I was wondering the same. I guess it makes sense to use a heavy weight model to make the entire design and split the work so that smaller models (possibly local one?) would then do the coding... But how would I even do that? I'm using Claude Code. Would I need support for this within the harness ?
I use Gemini 3 Flash, I've seen the Claude Code setups, bullish on Anthropic people are driving up tokens but I am able to produce outcomes with a fraction of the money.
Do you mind sharing your workflow? What do you mean by fraction of the money, in my case personally, I'm yet to reach a session limit on the subscription plan. I'm not "tokenmaxxing" as they say, so hard to see a scenario in which the plan is expensive for the value I get.
If you don't hit a limit running Opus, it means you are very much in the loop.
For example you probably don't have days where you ask Opus to review your whole code base and look for code duplication/technical debt/robustness issues, and then to fix some of the found issues, and do this 3-5 times until no big issues are found anymore.
What’s your prompt for this, the way you described it made it seem like there’s a generalizable way I can go about this. I just rely on a testing pipeline instead so can’t think of why I would need to proactively find holes where tests haven’t already done that for me.
It's so weird to me that the benchmarks remain so low, but the models are marketed as revolutionary. And if you say that low coding capabilities aren't a problem, say that to the token price hike and 'general use' model setup.
Why not sell it as a math agent? Why do I have to set up 4 agents to check each others' work?
Yeah the future is probably a number of highly specialised small models you can run on your own hardware rather than massive frontier models in the cloud.
MOE basically work that way already, QWEN/etc with low active params (A-number in name) allows to inference big models locally (only active params have to fit into memory)
That seems to be what Microsoft is betting on also based on what was shown at the BUILD keynote today + that new surface ultra and the surface mini PC with the new Nvidia chip. Nadella really played up local AI as the main use case they have in mind.
Please test your websites in Safari. Almost all of your iOS users use it by default, and the desktop experience is pretty close to the mobile experience, so testing is easy.
That scroll effect is jank city for me (yeah yeah works fine in Chrome/Edge).
Unless they specifically clarify that the testing and training benchmarks are completely separate, we have to assume they test on the same 'hill' the model climbs.
Please don't complain about tangential annoyances—e.g. article or website formats, name collisions, or back-button breakage. They're too common to be interesting.
I'd really like to get back to an autocomplete flow, ideally with some shared and optimized context with the relationship with my larger agent models.
But it seems like, by and large, even the faster models are now aimed at longer-running agentic flows and not sub-1s autocomplete. Or am I wrong about that?
is 51% good enough to reliably use? There's no world in which I use an AI agent where it gets even 15% of the code wrong, that's as bad a Tesla FSD where you need to pay attention to the road while engaging FSD. What's the point? My attention is what I'm trying to relieve, not mostly correct functionality. The only thing that matters is whether you can one-shot code like Claude or Codex, I'm not interested in a small but mostly-okay-but-annoyingly-buggy-every-now-and-then AI.
"Clean data" is impossible. Language models have polluted the landscape to such a degree it's impossible to filter them out now. OpenAI has no doubt discarded or muddled their dataset that was used to train the original ChatGPT, so there may be no dataset in existence now that isn't contaminated.
They're comparing to Haiku, not Opus. Haiku is currently at 4.5.
Even if it were Opus, comparing to a version number makes for an interesting snapshot of time comparison: if you knew how a model performed at whatever time in was in vogue, you can say "well, it looks like Model X is about 6 months/1 year/etc. behind the frontier SOTA" - which is exactly the discussion that happens in the open-weight/local LLM space. (interesting, MAI-Code-1-Flash does not appear to be such an open-weight model, following the western trend of locking models up)
Huh, according to that model card this is a 137B total parameter model.
Performance doesn't seem that good:
- MAI-Code-1-Flash (137B-A5B) = 51% on SWE-bench pro
- Qwen3.6-35B-A3B = 49.5% on SWE-bench pro (https://huggingface.co/Qwen/Qwen3.6-35B-A3B)
They benchmark against Claude Haiku but Haiku is not good, it's worse than tiny open models you can run locally or via API at 10% the cost.
The take away is that this model is a smaller model that competes with Haiku, I would hope they come out with a "Sonnet" competing model, then Opus. I have been wondering why Microsoft is kind of "sleeping" on offering models they themselves have made on Copilot, maybe it was part of their deal with OpenAI? Not sure.
They did release, MAI-Thinking-1 to compete with Sonnet. Totally not sure why that isn't at the top here.
Good question, and I missed that entirely!
Does anyone actually uses these smaller models for coding? If so, how? I usually Opus everything. Is the play to plan/design/architect with a heavier model than delegate structured tasks to these smaller ones? Would appreciate to hear someone's opinion on having done and tested both paths.
I am using Opus 4.x at work, and these "smaller" (20-80bn, 3-4bn active) models at home. Unfortunately there is no comparison, yet (IMHO anyway).
With Opus I can work, trust its designs, architecture suggestions, and code changes, even in a complex code base.
The smaller models seem to "try". They work for smaller tasks, but for more complex task it's often more work than doing it myself.
I wish it were different, and maybe in a year or two it will be.
I was wondering the same. I guess it makes sense to use a heavy weight model to make the entire design and split the work so that smaller models (possibly local one?) would then do the coding... But how would I even do that? I'm using Claude Code. Would I need support for this within the harness ?
I use Gemini 3 Flash, I've seen the Claude Code setups, bullish on Anthropic people are driving up tokens but I am able to produce outcomes with a fraction of the money.
Do you mind sharing your workflow? What do you mean by fraction of the money, in my case personally, I'm yet to reach a session limit on the subscription plan. I'm not "tokenmaxxing" as they say, so hard to see a scenario in which the plan is expensive for the value I get.
If you don't hit a limit running Opus, it means you are very much in the loop.
For example you probably don't have days where you ask Opus to review your whole code base and look for code duplication/technical debt/robustness issues, and then to fix some of the found issues, and do this 3-5 times until no big issues are found anymore.
What’s your prompt for this, the way you described it made it seem like there’s a generalizable way I can go about this. I just rely on a testing pipeline instead so can’t think of why I would need to proactively find holes where tests haven’t already done that for me.
plan using opus execute using local
It's so weird to me that the benchmarks remain so low, but the models are marketed as revolutionary. And if you say that low coding capabilities aren't a problem, say that to the token price hike and 'general use' model setup.
Why not sell it as a math agent? Why do I have to set up 4 agents to check each others' work?
It’s about bang for buck. That high a score for 5B params is pretty good, nigh unbelievable a short while ago.
It is my belief that smaller models will get better and better, and even cloud SOTA models will shrink.
Yet another reason the current buildout will feel like the railroads.
Yeah the future is probably a number of highly specialised small models you can run on your own hardware rather than massive frontier models in the cloud.
That's what I'm betting on anyway.
MOE basically work that way already, QWEN/etc with low active params (A-number in name) allows to inference big models locally (only active params have to fit into memory)
That seems to be what Microsoft is betting on also based on what was shown at the BUILD keynote today + that new surface ultra and the surface mini PC with the new Nvidia chip. Nadella really played up local AI as the main use case they have in mind.
The SOTA models will not shrink, because the problems will get bigger, from "write me a C compiler" to "clone Stripe business and run it".
The introductory blog post has a lot more information
https://microsoft.ai/news/introducingmai-code-1-flash/
and the model card
https://microsoft.ai/pdf/MAI-Code-1-Flash-Model-Card.PDF
The broader announcement of 7 MAI models seems to be where the 5B active in the title comes from
https://microsoft.ai/news/building-a-hillclimbing-machine-la...
Thanks! I've changed the top link to the blog post and put the other links in the toptext.
Shouldn’t the next model focus not be on code but system design?
Seems like the work from a good system design to code is practically solved.
Now it’s a matter of the design of the system. Or is that represented in these evals?
Have you tried system design with LLMs? I find them pretty good at suggesting 5 architectures for a problem and then iterating on the solutions.
Even if I had no idea, going with the default suggestion would not be a terrible mistake, assuming you did describe your requirements relatively well.
Please test your websites in Safari. Almost all of your iOS users use it by default, and the desktop experience is pretty close to the mobile experience, so testing is easy.
That scroll effect is jank city for me (yeah yeah works fine in Chrome/Edge).
not open weight or at least I did not find anything indicating open weight
So it's trained on the SWE Bench Pro evalset
What is your evidence for this claim?
They say hill climbing
https://microsoft.ai/news/building-a-hillclimbing-machine-la...
Unless they specifically clarify that the testing and training benchmarks are completely separate, we have to assume they test on the same 'hill' the model climbs.
Gemma 4 26B-A4B scored exceptionally well with 20% less params, so this isn't unprecedented.
You lost me at forced scrolling. Ugh!
From https://news.ycombinator.com/newsguidelines.html
Please don't complain about tangential annoyances—e.g. article or website formats, name collisions, or back-button breakage. They're too common to be interesting.
Would be cool if this were an open model.
To be clear about the size of the model: MAI-Code-1-Flash is 137B A5B.
It is good to se big companies like Microsoft launching LLMs. They have large amount of compute power and good scientists to create useful models.
Microsoft has been releasing LLMs for years.
They were mostly distilled or fine-tuned OAI models.
And occasionally un-releasing them like with WizardLM.
Sort of. Phi models were just trained on GPT outputs though.
"superintellegence team"
Why not assign them to make windows good :D
I'd love to see a tokens per second metric. I always prioritize speed over raw intelligence for flash models.
> I always prioritize speed over raw intelligence for flash models.
This model might have a perfect speed:
I'd really like to get back to an autocomplete flow, ideally with some shared and optimized context with the relationship with my larger agent models.
But it seems like, by and large, even the faster models are now aimed at longer-running agentic flows and not sub-1s autocomplete. Or am I wrong about that?
Scroll wheel hijacked on this entire domain
Fix:
Yeah this website is horrendous to use. What were they thinking?
You mean "what was the LLM thinking?"
is 51% good enough to reliably use? There's no world in which I use an AI agent where it gets even 15% of the code wrong, that's as bad a Tesla FSD where you need to pay attention to the road while engaging FSD. What's the point? My attention is what I'm trying to relieve, not mostly correct functionality. The only thing that matters is whether you can one-shot code like Claude or Codex, I'm not interested in a small but mostly-okay-but-annoyingly-buggy-every-now-and-then AI.
Claude opus 4.6 scores 51.9% on the same benchmark. Microsoft's result is quite good.
"Clean data" is impossible. Language models have polluted the landscape to such a degree it's impossible to filter them out now. OpenAI has no doubt discarded or muddled their dataset that was used to train the original ChatGPT, so there may be no dataset in existence now that isn't contaminated.
So it's not an open model while not being much better? Meh.
Comparing against Claude 4.5? Aren't we up to 4.8? But disingenuous?
They're comparing to Haiku, not Opus. Haiku is currently at 4.5.
Even if it were Opus, comparing to a version number makes for an interesting snapshot of time comparison: if you knew how a model performed at whatever time in was in vogue, you can say "well, it looks like Model X is about 6 months/1 year/etc. behind the frontier SOTA" - which is exactly the discussion that happens in the open-weight/local LLM space. (interesting, MAI-Code-1-Flash does not appear to be such an open-weight model, following the western trend of locking models up)
Latest Haiku (smallest Anthropic Model) is version 4.5, they haven't released a new version, hence the comparison to that.