Composer is a new model we built at Cursor. We used RL to train a big MoE model to be really good at real-world coding, and also very fast. cursor.com/blog/composer Excited for the potential of building specialized models to help in critical domains.

Oct 29, 2025 · 4:32 PM UTC

53
74
43
786
Our benchmark comparisons of the model.
5
2
5
54
One personal reflection is how interesting a challenge RL is. Unlike other ML systems, you can't abstract much from the full-scale system. Roughly, we co-designed this project and Cursor together in order to allow running the agent at the necessary scale.
1
7
69
Replying to @srush_nlp
Nice model @srush_nlp you've been cooking! 😃
1
3
Thanks! Although I am an IC these days on a stacked team. RL dev definitely a team sport. Just did my little part.
1
13
Replying to @srush_nlp
Was this what was once Cheetah? If so Composer is not available in cmd+k inline edits and that was my favorite place to use Cheetah.
2
5
Cheetah was the v0 of this model primarily to test speed. > In our development process, we experimented with a prototype agent model, codenamed Cheetah, to better understand the impact of faster agent models. Ah I will tell the product team.
2
2
19
more replies
Replying to @srush_nlp
Is the speed owing to the inference or the architecture?
1
1
Some details on speed here: cursor.com/blog/kernels Architecture: It's a large-scale mixture-of-experts model, so what you might expect.

1.5x faster MoE training with custom MXFP8 kernels · Cursor

Built to make you extraordinarily productive, Cursor is the best way to code with AI.

cursor.com
2
1
1
17
more replies
Replying to @srush_nlp
Do you think text diffusion models still have future in coding tasks?
1
A lot of very smart people are working on it. Watching closely and hoping they crack some of the challenges.
1
Replying to @srush_nlp
So how does the model perform compared to gpt 5 etc ? I don't get this chart....
1
It performs worse than GPT-5 but runs much faster. GPT-5 is a very smart model.
1
2
7
more replies
Replying to @srush_nlp
What is it's performance across languages? Is it optimized for web-development or also for backend, low-level, etc.?
1
A key criteria in the benchmark was to be diverse across languages, domains, location on the stack. Our main target though was realistic, large codebase settings (as opposed to very hard, algorithmic problems).
1
2
more replies
Replying to @srush_nlp
thanks for all the hard work @srush_nlp & rest of ml team!
1
8
Replying to @srush_nlp
This is incredible! I have been curious about this for long: Does cursor get to train on the diffs made to code by frontier models? Or is that considered a no-go because of being a derivative of the outputs of gpt/claude?
5
Replying to @srush_nlp
congrats! 🎉
3
Replying to @srush_nlp
imagine how much faster it’d be on groq 🥹
2
Replying to @srush_nlp
So, Cheerah was indeed a model by @cursor_ai! I found it changes the way I use AI agents because it is so fast that I can stay in the loop when working with it. Seems Cursor is capitalizing on the huge amount of data they have from all people using their IDE
2
Replying to @srush_nlp
ai had its “think longer” moment. feels like we’re headed into an era of domain specific RL and “work fastest” as the benchmark 80% of sonnet 4.5 quality at 8x the speed is more valuable than 100% the quality at 1/8th the speed i would expect many companies to follow (where feasible)
1
Replying to @srush_nlp
crazy that cursor is so popular + good despite never having trained their own full model until now. takeaway: there is a ton of value to be made of off pre-exisiting open source models + making them fast. You can get pretty far just by post-training a lot
Replying to @srush_nlp
Kudos, Sasha! :) 👏
Replying to @srush_nlp
let me use the model directly!
gary IH fung
 @garyfung
Oct 29
as an ex-Cursor user, let me preamble that I'm looking for reason to come back but this is another chart crime - what's best open and frontier? Tell us how you are cherry picking! - intelligence: Cursor-Bench score. That's like me saying I'm the absolute best according to GaryScore why are you like this. Give me legit reasons to come back from @zeddotdev and Codex!
Replying to @srush_nlp
Incredible work. Composer sounds like a major leap toward faster, more reliable AI-assisted coding.
Replying to @srush_nlp
What is the MoE - Qwen?
6
Replying to @srush_nlp
Would you guys disclose which open weight model you guys trained on?
1
3
Replying to @srush_nlp
It’s so fast, y’all killed it
1
Load more