Ornith-1.0 introduces open-weight models for agentic coding, built on Gemma 4 and Qwen 3.5

DeepReinforce has released Ornith-1.0, an open-weight model family under the MIT license, marking the organization’s first model release. The variants include 9B Dense, 31B Dense, 35B MoE, and 397B MoE, all built on pretrained Gemma 4 and Qwen 3.5 base models.

According to the announcement, Ornith-1.0 achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks. The post notes compatibility with the Apache 2.0 licenses of the underlying Gemma 4 and Qwen 3.5 models, which the author states are not encumbered by additional restrictions present in prior Gemma terms.

The author provides practical observations from running the 35B MoE variant via LM Studio using the ornith-1.0-35b-Q4_K_M.gguf quantized file (20GB). They report that the model proficiently handles multi-step agent harness tool calls, including locating specific code in a Datasette repository and generating a simple illustration of a pelican at a rate of 103 tokens per second.

The post also highlights limited public information about DeepReinforce, noting the earliest identified work from the group is a June 2025 paper titled 'CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning.'