Researchers who want an open, MIT-licensed coding agent to study or fine-tune; the team's own figure is 42.2% on SWE-Bench Verified.
Kenning's take
What it does
DeepSWE-Preview is an open coding agent model from the Agentica team and Together AI, trained only with reinforcement learning on 4,500 real software tasks. The team reports 42.2% pass@1 on SWE-Bench Verified, rising to 59% with test-time scaling. Weights are on Hugging Face under the MIT licence for people to run themselves.
- Trained from Qwen3-32B with reinforcement learning only
- Open weights under the MIT licence
- Built with the open-source rLLM framework
- 42.2% pass@1 on SWE-Bench Verified, per the authors
Who it suits. Researchers and developers who want an open coding agent model to run, study or fine-tune.
Key facts
- Category
- AI models
- Made by
- Agentica and Together AI
- Released
- 1 July 2025
- How to get it
- free download of weights on Hugging Face (agentica-org/DeepSWE-Preview)
- Works on
- open weights
- In the UK
- Yes
- Free plan
- open weights, MIT licence
FAQ
Is DeepSWE free?
Yes. The weights are published on Hugging Face under the MIT licence.
What model is DeepSWE based on?
It was trained from Qwen3-32B using reinforcement learning only.
When was DeepSWE released?
The Hugging Face model page was created on 1 July 2025.
Facts from the maker's own pages and the sources listed. Checked 2026-10-11.
Similar tools
ACE-Step 1.5
Open-source music generation model that writes full songs with lyrics and runs on ordinary consumer GPUs.
AlphaGenome Atlas
Google DeepMind database of AI-predicted effects for every possible single-letter change in human DNA.
Amazon Nova 2
Amazon's second family of Nova AI models, covering reasoning, speech conversation and combined text-and-image generation.
Cisco Antares
Cisco's small open-weight models that find where known security vulnerabilities sit in a codebase, running locally.