Ornith-1.0-35B

About

Ornith-1.0-35B (GGUF builds at deepreinforce-ai/Ornith-1.0-35B; card at ornith-ai/Ornith-1.0-35B) is the lightweight 35B-MoE member of Ornith 1.0, the self-improving family of open-source models for agentic coding from deepreinforce-ai, designed for efficient single-GPU deployment. Ornith 1.0 ships in 9B-Dense, 31B-Dense, 35B-MoE and 397B-MoE variants post-trained on top of Gemma 4 and Qwen 3.5, achieving state-of-the-art performance among open-source models of comparable size on coding benchmarks such as Terminal-Bench 2.1, SWE-Bench, NL2Repo and OpenClaw.

Its signature is the self-improving training framework: RL teaches the model to generate not only solution rollouts but also the scaffolds that drive them - jointly optimizing scaffold-rollout co-optimization so the model discovers better search trajectories and higher-quality solutions. This record's lineage: post-trained with RL (self-scaffolding) on top of the Qwen3.5 35B base. The underlying model is a Qwen3.5-style sparse Mixture-of-Experts with 35B total and 3B active parameters: 40 layers mixing 30 linear-attention layers (Gated DeltaNet-style) with 10 full-attention layers (interval 4), 256 routed experts with 8 active per token plus 1 shared expert, one Multi-Token Prediction layer, 248K vocabulary and a 262,144-token context, with an optional Qwen3.5 ViT for image input on the multimodal builds.

Serving is optimized for GGUF runtimes: llama.cpp, Ollama, LM Studio and vLLM/SGLang OpenAI-compatible endpoints, with agent frameworks (Hermes, OpenClaw, OpenHands, opencode) and Unsloth for local fine-tuning. Released June 21, 2026 under the MIT License.

Training Data Post-trained (RL with self-scaffolding) on top of Qwen 3.5 35B base. RL jointly optimizes scaffold and solution rollouts for agentic coding.