In brief: Ornith-1.0 offers agent-driven capabilities for code tasks in sizes 9B, 31B, 35B MoE, and 397B MoE, achieving state-of-the-art performance on coding benchmarks at comparable scale.
The new Ornith-1.0 model from DeepReinforce is available under the MIT license and is designed to control agent tasks in programming. It is provided in four variants and is based on Gemma 4 and Qwen 3.5.
DeepReinforce has released Ornith-1.0, a new open model comprising four variants: 9B Dense, 31B Dense, 35B MoE, and 397B MoE. The model is based on the pretrained models Gemma 4 (Apache 2.0 license) and Qwen 3.5 (Apache 2.0 license), whose licenses are compatible with the release.
In practical tests, Ornith-1.0 demonstrates robust execution of agent tasks with multiple tool calls in sequence. In a test session with a Datasette repository, the model executed complex code search queries—such as identifying actor cookie decoding and locating the code section that opens an insert dialog. With the 35B MoE variant (GGUF format, 20 GB) in LM Studio, the model achieved approximately 103 tokens per second on text rendering.
So far, little is known about DeepReinforce itself. The earliest traceable publication from the team is the paper “CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning” from June 2025, which addresses optimizations for GPU computing.
Source: simonwillison.net · Published June 29, 2026
Lumi AI News — AI-assisted curation pursuant to Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.2.