DeepReinforce has released Ornith-1.0, a family of open-weights language models distributed under the MIT license and designed for agent-style coding workflows. The release includes 9 billion and 31 billion parameter Dense models, as well as 35 billion and 397 billion parameter Mixture-of-Experts (MoE) variants.
Foundations and performance
Ornith-1.0 is built on top of existing pretrained models: Gemma 4 and Qwen 3.5. According to early reports, the family achieves state-of-the-art results among open-source models of similar size on coding benchmarks.
Both Gemma 4 and Qwen 3.5 are released under the Apache 2.0 license, which the report notes is compatible with building Ornith-1.0 on top of them. The Gemma 4 release in particular is not subject to the additional Gemma Terms of Use that affected earlier Gemma models.
Practical usage and deployment
A user report describes running the Ornith-1.0-35b model packaged as a GGUF file named "ornith-1.0-35b-Q4_K_M.gguf", approximately 20 GB in size. The model was executed via LM Studio and connected to a local device referred to as "Pi." Initial impressions were positive: the model handled agent harnesses and multi-tool call workflows proficiently.
In a terminal session the user asked the model to "find the code that decodes the actor cookie" and to "find the code that opens the insert dialog when the button is clicked" against a Datasette checkout; the model completed these tasks with ease.
The same setup was also used to have the model "draw a pelican," with generation running at 103 tokens per second. The resulting image was somewhat mangled but recognizable as a pelican.
About DeepReinforce and publications
The report found little public information about DeepReinforce itself. The earliest paper located under the group's name is "CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning," dated June 2025.
Why this matters
Ornith-1.0 is notable because it provides entirely open-weight models, permissively licensed, that are tuned for agentic coding tasks and made available in multiple sizes including large MoE configurations. The availability of a 20 GB GGUF pack and LM Studio compatibility lowers the barrier for local deployment and practical experimentation by developers.
Source details (from the report)
- Model family: Ornith-1.0 (DeepReinforce)
- License: MIT (model family)
- Variants: 9B Dense, 31B Dense, 35B MoE, 397B MoE
- Base models: Gemma 4 (Apache 2.0), Qwen 3.5 (Apache 2.0)
- Available GGUF: ornith-1.0-35b-Q4_K_M.gguf (~20 GB)
- Local testing: LM Studio + Pi; generation example: 103 tokens/s
- Earliest related publication: "CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning" (June 2025)
The facts above are drawn from the model announcement and early user reports.
