Ornith-1.0 wants to close the coding gap, for free
DeepReinforce just dropped Ornith-1.0, a family of open coding models that don't just generate solutions, they use reinforcement learning to improve the reasoning scaffolds that guide those solutions. Four sizes, from a 9B dense model up to a 397B MoE, all built on Gemma 4 and Qwen 3.5 bases, all MIT licensed with weights on Hugging Face in bf16, FP8 and GGUF.
The numbers are the story. The 397B model hits 82.4 on SWE-bench Verified against Claude Opus 4.8's 87.6, and 78.2 on Terminal-Bench 2.1 versus 78.9. That's an open, free-to-run model landing within a few points of the best proprietary coding agent on the planet, with a 256K context window and native tool-calling.
The deeper idea is what makes this part of a pattern. For weeks now the interesting work has been models that improve themselves: SIA updating weights and harness in one loop, MLEvolve out-evolving AlphaEvolve, Anthropic reporting Claude writes most of its own code. Ornith puts that loop inside an open model anyone can download, optimizing both the answer and the thinking framework that produced it.
If you've been waiting for an open coding model you can actually run in production without renting frontier API calls, this is the closest anyone has gotten. The gap to Opus is now small enough to matter. Link: github.com/deepreinforce-ai/Ornith-1
← Back to all articles
The numbers are the story. The 397B model hits 82.4 on SWE-bench Verified against Claude Opus 4.8's 87.6, and 78.2 on Terminal-Bench 2.1 versus 78.9. That's an open, free-to-run model landing within a few points of the best proprietary coding agent on the planet, with a 256K context window and native tool-calling.
The deeper idea is what makes this part of a pattern. For weeks now the interesting work has been models that improve themselves: SIA updating weights and harness in one loop, MLEvolve out-evolving AlphaEvolve, Anthropic reporting Claude writes most of its own code. Ornith puts that loop inside an open model anyone can download, optimizing both the answer and the thinking framework that produced it.
If you've been waiting for an open coding model you can actually run in production without renting frontier API calls, this is the closest anyone has gotten. The gap to Opus is now small enough to matter. Link: github.com/deepreinforce-ai/Ornith-1
Comments