Ornith-1.5: MIT-licensed open LLMs that match Opus on coding
Three sizes from 9B to 397B, weights on Hugging Face, and a loop that writes its own training tasks — the 9B posts 70.6 on SWE-bench Verified.
Copy markdown
Three open sizes, one MIT license
Ornith-1.5 ships as a 397B mixture-of-experts flagship, a 35B MoE with ~3B active per token, and a 9B dense model — all open-weight under MIT, with a 256K context window and weights already on Hugging Face. Fine-tune or ship them commercially with no strings.
The 9B runs on your laptop tonight
The 9B dense model scores 70.6 on SWE-bench Verified and ~47 on Terminal-Bench 2.1 — matching much larger models like Qwen 3.6-35B — and ships with a quantized build for iPhone and Android. Pull it and wire up a local coding agent this weekend.
It writes its own training tasks
The 'self-improvement' loop has the model propose progressively harder tasks, build the scaffold — tools, decomposition, orchestration — and generate solution rollouts, then reward all three jointly with GRPO. It's an open model that levels up without anyone hand-labeling the next dataset.
How close to Claude Opus, really
The 397B flagship hits 86 on SWE-bench Verified and ~85 on Terminal-Bench 2.1 — edging Claude Opus 4.8 — plus 92.8 on GPQA Diamond and 86.6 on BrowseComp. It trails Opus on DeepSWE (56 vs 59), so it trades blows rather than sweeps, but it's the closest open weights have come.