DeepReinforce has released Ornith-1.0, an open weights model under the MIT licence that focuses on agentic coding tasks. The team offers variants ranging from a 9B dense model up to a massive 397B mixture of experts. These versions build upon pretrained weights from Gemma 4 and Qwen 3.5, both of which carry Apache 2.0 licences that allow this specific usage. Benchmarks show the new model achieves state-of-the-art performance among open-source options of comparable size for coding challenges.
Testing with LM Studio and the Pi interface suggests the system handles multi-step tool calls proficiently. A terminal session demonstrated the model locating specific code within a Datasette checkout without issue. Another test involved drawing a cartoon pelican riding a bicycle, which the model generated at 103 tokens per second. The resulting image was recognisably a pelican despite some minor distortions. Information about the company itself remains scarce, with their earliest published paper appearing in June 2025.
* Available in 9B, 31B, 35B, and 397B parameter sizes
* Built on Gemma 4 and Qwen 3.5 architectures
* Licensed under MIT with Apache 2.0 underlying weights
