Introducing DeepCoder-14B-Preview - our fully open-sourced reasoning model reaching o1 and o3-mini level on coding and math. The best part is, we’re releasing everything: not just the model, but the dataset, code, and training recipe—so you can train it yourself!🔥 Links below:

Excited to present our project in collaboration with Agentica: 14B LLM trained with Code RL that reaches OpenAI's o3-mini-low performance on coding benchmarks like LiveCodeBench, Codeforces and HumanEval! We open source the code, data, weights and full recipe

ollama run deepcoder 🫡 let’s go open source!!

Nice!

New article from VentureBeat about our model!

I’ve seen some people using DeepCoder-1.5B as speculator for DeepCoder-14B. Because the final stage for both was “self-play” RL, it won’t be a good speculator, as they diverge a lot. If there’s enough interest, we can train a good speculator (w/ logit distillation) & release

Our DeepCoder-14B LiveCodeBench v5 scores have been validated and put on the official leaderboard!

DeepCoder has reached the top of HuggingFace trending models 🥳

Excited to join some great friends at Nous as one of the judges for their first hackathon! It will be focused on RL environments. Pull up, it will be fun :)

🚀 Introducing DeepSWE 🤖: our fully open-sourced, SOTA software engineering agent trained purely with RL on top of Qwen3-32B. DeepSWE achieves 59% on SWEBench-Verified with test-time scaling (and 42.2% Pass@1), topping the SWEBench leaderboard for open-weight models. 💪DeepSWE

Excited to introduce DeepSWE-Preview, our latest model trained in collaboration with Agentica Project Using only RL, we increase the performance of Qwen 3 32B from 23% to 42.2% on SWE-Bench Verified!

Soham Parekh was a DeepSWE checkpoint sorry

It's easy to confuse Best@K vs Pass@K—and we've seen some misconceptions about our results. Our 59% on SWEBench-Verified is Pass@1 with Best@16, not Pass@8/16. Our Pass@8/16 is 67%/71%. So how did we achieve this? DeepSWE generates N candidate solutions. Then, another LLM

Let’s normalize reading through and actually understanding something before attempting to criticize it publicly :)

important correction on DeepSWE-Preview. on SWE-Bench-Verified: Pass@1 = 42.2% "Best@8" = 59%, trajectory selection achieved with a hybrid (execution-free + test-based, as in R2E-Gym paper) verifier. Ie the system itself can yield 59% w/o ground truth check. Actual pass@8 =67%.

Only after labelling a dataset by your own you know how dirty it is