MAI-Code-1-Flash is now available for GitHub Copilot

[ebndev]

MAI‑Code‑1‑Flash, Microsoft's latest small‑tier coding model, is now rolling out in GitHub Copilot starting with VS Code. MAI‑Code‑1‑Flash delivers best‑in‑class quality for its size, outperforming other small models in early testing, making it well‑suited for lightweight coding workflows. Designed and tuned specifically for GitHub Copilot, this marks the first in a new wave of purpose‑built coding models from Microsoft.

Availability in GitHub Copilot

MAI‑Code‑1‑Flash is beginning to roll out to Copilot Free, Pro, Pro+, and Max plans. Availability will start with a limited set of users and expand gradually over the coming weeks.

You'll be able to select the model in the model picker in

• Visual Studio Code

Learn more

To learn more about the new model, see the MAI-Code-1-Flash Model Page.

To explore all models available in GitHub Copilot, see our documentation on models and get started with Copilot.

Leave a comment

Join the discussion and leave feedback in the comments below!

For all releases announced during Microsoft Build, check out our Community Recap discussion 🌟

[Jaimi]

More than interested to try when it comes out. But what about the full fledged Thinking version?

[yanancai]

We are actively working on that, please stay tuned!

[hypestwave]

thanksx

[natestrom]

Definitely looking forward for the new MAI models to land on Business/Enterprise accounts!

[yanancai]

Thanks for your interest, we are actively working on a plan to enable preview for Enterprise/Business customers. Will share more once we get a more concrete process of onboarding!

[ctolkien]

+1 for this... given the recent pricing changes, getting token efficient models is now key for Copilot usage.

[ederbond]

Is it going to be available to download and run completely offline on my own hardware with the full fledged Visual Studio 2026 IDE?

If they don't want to lose, even more developers from their ecosystem other platforms, they should either release it public to be downloaded on hugging face to allow us to run it locally on the developer's machine or at least release a full version of it built-in into VS 2026 able to run local and for free. Another option could be a variant of this model that allow at least full agentic programming with .NET only, with the goal do run it locally and Free (at least for .NET) it will be the only effective way to contain the competition to defeat you) you're otherwise the competition might defeat you.

[yanancai]

Thanks for your feedback @ederbond! We are working on our overall open source strategy, so super helpful to hear your desired use cases here! May I ask what kind of hardware you have? Trying to get a sense of the memory footprints.

[ederbond]

@yanancai I have a Dell Mobile Workstation with

• 64GB RAM +
• Intel Core Ultra 9 285H (With an NPU) + an integrated Intel Arch GPU
• Dedicated NVidia RTX 5070 with 8GB VRAM

Which I'm planning to replace by a Surface Laptop Ultra with that amazing NVidia Arm CPU when it gets launched by the end of this year (if the price is reasonably affordable).

[ederbond]

@yanancai talk with @madskristensen (the lead of Visual Studio 2026 team) about the huge benefit it could bring to Microsoft's flagship IDE if these MAI models could be infused inside Visual Studio 2026 to run this family o MAI models, locally in agentic coding mode.
If Microsoft could make it possible, a lot of developers (even from other languages) will jump in, into Visual Studio and the product will become way more attractive if Developers could run AI locally.

[ederbond]

It will help selling more Visual Studio subscriptions, + Windows licenses, + Azure + Surface Hardware and make the entire Microsoft's ecosystem stronger.

[gimlichael]

Sadly, VS 2026 is down prioritized from Microsoft 😞

[francovp]

Why no business or Enterprise? :(

[yanancai]

Thanks for your interest, we are actively working on a plan to enable preview for Enterprise/Business customers. Will share more once we get a more concrete process of onboarding!

[spoiledsport]

WHy not Visual Studio?

[drewid]

100%. About to dump tooling due to the 20x-30x increase in costs with GitHub Copilot and this is maybe the only thing that would save it.

[yanancai]

Feedback taken, thanks for sharing this, it helps us prioritize the clients support.

[okasi]

Will it be uploaded to hugging face?

[yanancai]

This model is a close-weight one. We are working on our overall open source strategy, would be interested in hearing your desired use cases here!

[mindplay-dk]

Put it up on OpenRouter so we can try it, please - and so we can see how much it really costs.

[OminousIndustries]

Put it up on OpenRouter so we can try it, please - and so we can see how much it really costs.

Agree with this. As far as releases go this one is making it rather difficult to even try the model

[mindplay-dk]

They have the other MAI models (image, voice, transcribe) already up there, so presumably it's coming.

[yanancai]

Thanks for the feedback @mindplay-dk, just to confirm, would you like to use the model via OpenRouter using clients like VS Code? Or connects with other coding agent clients?

[chriddyp]

This is looking great! Would love to try it out and benchmark it in Plotly Studio for data visualization and analytics workflows. Will this be available as an API on Azure or Open Router? Or only integrated in product suites?

[yanancai]

Thanks for your interest, we are actively working on a plan to enable preview for Enterprise/Business customers. Will share more once we get a more concrete process of onboarding!

[winter123123]

Why are business/enterprise users not supported?
Costs have already increased due to the shift to a pay-as-you-go model, so why do we have to face this kind of discrimination?
If this is how we are treated, we will stop using the enterprise service.

[drewid]

When can we get this in regular Visual Studio? About to dump tooling due to the 20x-30x increase in costs with GitHub Copilot and this is maybe the only thing that would save it.

[tfukumorikkmic]

I agree.

[yanancai]

Feedback taken, thanks for sharing this, it helps us prioritize the clients support.

[gimlichael]

Eager to try this - I need a replacement for Claude Haiku 4.5 - and this looks like the sweet spot. When will it be available? I am on the Pro+ and already 2 day has passed 🐌

[yanancai]

Thanks for your interest! We are still rolling out to model but it should be available to all individual users soon. Thanks for your patience.

[gimlichael]

I had high hopes for this model - especially looking at your benchmark. Sadly, Haiku 4.5 excel compared to this. Hope it will get better though 😞

[erincon01]

I don't see MAI in my ghcopilot business. Is this expected?

[MihailCosmin]

I also don't see it as of 08 June
Is this an US, EU rollout issue ?

[yanancai]

Thanks for your interest, we are still gradually rolling out the model to all users and are working on a plan to enable preview for Enterprise/Business customers. Will share more once we get a more concrete process of onboarding!

[eugenenell2021]

When will MAI be available for Enterprise GitHub Copilot? Why are business/enterprise users not supported as yet and how long will it take before we get it?

[yanancai]

Thanks for your interest, we are still gradually rolling out the model to all users and are working on a plan to enable preview for Enterprise/Business customers. Will share more once we get a more concrete process of onboarding!

[oduis]

Could it be that "available" just means "for inline code completion", but if never shows up in the models?

[gimlichael]

No - it's literally written that it start with a limited set of users and expand gradually over the coming weeks. Go have a look at YouTube - people lucky enough to get the model are quite pleased with it. And i actually think that Copilot will have value again as I assume Microsoft can provide these models cheap (since they own everything tied to it).

[oduis]

Thanks, just watched the Youtube demo.
Unfortunately, the price is exactly the same as GPT5.4 mini: GitHub Model prices
From the SWE Pro bench MAI-Code-1 and GPT5.4 mini seems to be in the same ballpark, and I found the latter really usable for its size.
I would have hoped for a lower cost one like the Microsoft Raptor model though.

[gimlichael]

At least for me, it looks like a solid replacement for Claude Haiku 4.5 - but time will show 🤞🏼

[yanancai]

Thanks for your interest! We are still gradually rolling out the model to all users. And noted for your feedback on pricing.

[tanvishinde017]

Thanks for bringing MAI-Code-1-Flash to GitHub Copilot.

It's great to see more purpose-built coding models being added to the ecosystem. Having efficient lightweight models is especially valuable for quick edits, smaller projects, and reducing overall credit usage.

Looking forward to trying it out in VS Code and seeing how it compares with the other available models. It would also be interesting to see future benchmarks and broader availability across additional Copilot surfaces.

[justinoneten]

I'd love to hear the line of thinking as to why this was rolled out to individual accounts first vs. business / enterprise. Shouldn't users paying for a premium product have the earliest access?

[Jaimi]

I see what you mean. But look at it this way - if it goes catastrophically bad, it's better for them to test it on those accounts.

[ederbond]

Microsoft's free plan in VS Code, its terribly small. Tested it on a small ASP.NET WebAPI app the it's consumed 23% in a single small change of creating a db migration in a small project. It was in High thinking mode. It will be like a 1 ~ 3 days trial.

[ederbond]

If Microsoft don't want to lose even more DEVELOPERS from their ecosystem to other platforms, they should either release it on hugging face to allow us to run it locally on our machines or at least release a full version of it built-in right into VS 2026 that's capable to run local and for free.

Another option could be a variant of this model that allow at least full agentic programming with .NET only, with the goal do run it locally and Free (at least for .NET) it will be the only effective way to stop the competing AI players to defeat you, otherwise .NET might gonna die and by consequence hit Windows harder.
Cause as you might have learned from the age of Window Phone:

1. No Developer == No App
2. No App == No Consumers OS
3. No Consumer on OS == Return to number 1; and keep on this endless dead loop.

In the meanwhile Google and Apple are welcoming the entire dev community to their tools with their own models.

[yanancai]

Thanks for the feedback! We are working on a plan to enable preview for Business/Enterprise users, please stay tuned.

[justinoneten]

@yanancai I'm sure it's in the roadmap, but my question is why paying users weren't prioritized over those who use it for free.

[PeterLeeSAG]

May I know how to apply or select it in the agent mode on the web UI? I can't see it on the page of the agent mode.

[yanancai]

The model is still being rolled out gradually, first in VS Code, then to CLI and other clients and surfaces. It is not yet available on the web UI at this moment. Thanks for your interest and patience!

[doggy8088]

Does MAI-Code-1-Flash support vision capabilities?

[yanancai]

Not yet, but coming soon.

[dehlers-cts]

Still not available in GitHub Copilot Enterprise. It was released on June 2 but still not available. What is up Microsoft?

[doggy8088]

https://x.com/MicrosoftAI/status/2065133021049782491

[blkst8]

I updated the GitHub Copilot extension in my IDE and VS Code, but I can only see the new model in the VS Code extension.

Also, the new model is NOT good even for small tasks. It makes a lot of mistakes.

[gimlichael]

In the announcement they outline its only available in VSCode .. weird, but communicated ..

[kpobb1989]

The model is good, but it does not support images, which makes it useless when working in VS Code and trying to use Playwright tool to adjust frontend changes. That's why I still prefer Haiku.

[tstooke]

I've finally gotten to play with the new model for a few days, so I wanted to drop in my initial feedback.

Given the token efficiency and the speed, this thing is AMAZING. It does still make some mistakes and skip some tasks if the prompt is too complex or the session is too long; however, even the most capable (and expensive) frontier models do that sometimes. I've kept my sessions with it relatively short and focused, and I haven't seen any issues with it going completely off the rails. The generated code has been quite accurate, it follows instructions very well, and it can even handle codebase investigations and assessments well. It has easily rivaled GPT-5.3-Codex, which is often the model selected by "auto" mode, and it doesn't end every response with "if you want me to, I can..." I would also say that it would be easy to recommend this MAI model over Sonnet 4.6 for many tasks, given the cost differential.

These are the two big things that I'm really looking forward for this model:

1. Availability in Enterprise/Business accounts. This has the potential to really help us reign in the token usage and still get some great work done.
2. Image support. An important part of the work that we do involves providing screenshots to the model for analysis and assessment (similar to the previous comment about the Playwright workflow).

So, at this point, it's not going to completely replace any other model; however, it shows some really great potential, especially now that token usage is such a huge factor in determining AI value.

Keep up the great work, @yanancai and team!

[Randalphwa]

I've found that models works best if you use the model to create custom agents -- the way instructions are worded can make a big difference to how well a model works. After having Mai create a custom subagent for editing C++23 code, I had a frontier model look at a fairly complex C++ project of mine and come up with three moderately difficult tasks for agents to create new src/hdr files that required reading and understanding one or more class header files in the project. I use DeepSeek V4 Flash with high thinking as my workhorse since it outperforms Sonnet 4.6 that I used to use. So, I had my frontier model run a V4 Flash thinking model and Mai-Code-1-Flash model with the identical task using each of the custom sub agents written specifically for the model they ran on.

First run, Mai didn't use C++23 standards as instructed, nor did it follow coding practices in the rest of the project. I made some minor tweaks to the custom model instructions, and ran the tasks again and those problems went away. On one of the tasks, the header files were byte-for-byte exactly the same. The logic was identical in all three tasks. Mai was slightly more conservative about pointers, so it wrote a bit more code to prevent using potentially null pointers (in this case the pointers could never be null, but I appreciated the extra caution anyway).

Watching Mai work with the Agent Debug Log up is quite impressive at how fast it calls and processes tools. It is still more expensive to user per task then DeepSeek V4 Flash as long as you don't count your time, or you are able to switch off to a different project so that you can keep working. The 264K context window means it won't run as long as V4 Flash with it's 1 million context window, but as an everyday workhorse, it can certainly keep up with Sonnet, V4 Flash, and GPT-4.1 -- at least in my experience so far.

A caveat to this review is that while throwing some moderately difficult real-world tasks at it is a valid comparison, I have yet to use it enough to truly understand where it's weaknesses are. I also have no idea how well it works with standard VSCode agents versus my highly customized agents -- so that might also be why it's working well for me, and might not work as well for people just using copilot out of the box with no customization.