Ornith-1.0 generates its own agent orchestration — no fixed harness required
DeepReinforce's Ornith-1.0 (9B–397B MoE, MIT licensed, built on Gemma 4 and Qwen 3.5) learns to generate task-specific orchestration logic via reinforcement learning rather than operating inside a fixed human-designed agent harness — it claims SOTA among open-source models of comparable size on coding benchmarks. The 35B MoE GGUF (20GB) runs in LM Studio at 103 tokens/second, making a self-orchestrating coding agent runnable on a single consumer GPU today.
Source: simonwillison.net ↗
it seems to be able to run the agent harness over many tool calls in a proficient way
Simon Willison
Why this matters
- → Open-source model runs self-orchestrating agents on single consumer GPU today
- → Eliminates need for fixed agent harness; learns task-specific logic via RL
- → SOTA coding performance among comparable-size open models, MIT licensed
Self-scaffolding agents