Ornith-1.5: open MIT coding models that rival Opus 4.8, run local
DeepReinforce ships three sizes under MIT: a 397B that ties Opus 4.8 on SWE-bench, a 3B-active 35B, and a 9B for your phone. Weights, GGUF and MLX are live.
Copy markdown
The 397B ties Claude Opus 4.8 — under MIT
The flagship Ornith-1.5-397B MoE posts 86.0 on SWE-bench Verified, 86.1 on Terminal-Bench 2.1 and 92.8 on GPQA Diamond — level with or ahead of Claude Opus 4.8, but downloadable and self-hostable under the MIT license. No API meter, no usage terms on the weights.
The 35B-A3B is the one you'll actually run
Only ~3B of its 36B parameters fire per token, so the mid-size MoE stays fast on a single 24-48GB GPU while hitting 79% on SWE-bench Verified. bartowski and unsloth GGUF quants are already trending on Hugging Face — near-frontier coding on hardware you own.
A 9B that runs on your phone
The dense Ornith-1.5-9B ships iOS and Android builds and beats the larger Gemma-4-31B on most tests — enough to embed an offline coding or agent assistant on a laptop, edge box, or handset with zero network round-trips.
Drop it into your agent stack today
262K context (~1M with YaRN), weights in BF16/FP8/NVFP4/GGUF/MLX, and an Ollama build (ornith-claude-coder) that already serves it behind Claude Code. Point Cline, Continue or Claude Code at a local endpoint and keep your source off someone else's servers.
Why a 3B-active model codes this well
Ornith trains with a self-scaffolding loop: the model invents progressively harder tasks, attempts them, and reinforces on validity, novelty and difficulty (aiming for a ~0.2 solve rate). It's the recipe DeepReinforce debuted with June's Ornith-1.0, now scaled into a full open family.