Codeset

Try it now at codeset.ai 🎁Free run for the first 50 users with code CODESETLAUNCH

Results? On our benchmarks, Haiku + Codeset performs better than vanilla Opus at nearly 1/10th the cost! The next leap in AI coding won't come from smarter models alone. It'll come from models that finally know where they are.

We made Haiku as good as Opus. Without changing the model. Today we're launching Codeset. Here's what we found 🧵

New blogpost is out! Read it at codeset.ai/blog/building-…

Try it now: 🔗 Platform: codeset.ai 📚 Docs: docs.codeset.ai 💾 Dataset: huggingface.co/datasets/codes… $5 free credits, pay-as-you-go, no minimum commitment. We'd love your feedback as we build this out.

What's next: more languages, additional task types beyond issue resolution, support for other public datasets, and custom datasets on commission. We're building this for teams working on the next generation of code agents.

We're launching Codeset today 🚀

Read more about CWM at ai.meta.com/research/publi… Read more about Codeset at codeset.ai

We believe that the future of code LLMs lies in understanding computational environments. For this, we need to scale executable data and improve both state observability and verification. Our platform will enable anyone to easily train models that reason about code execution.

🔎 Paper 2 in the Codeset Paper Series: CWM - An Open-Weights LLM for Research on Code Generation with World Models by @TacoCohen @adiyossLC @syhw and the Meta FAIR CodeGen Team