Ornith-1.5: MIT open weights from a 397B MoE down to a phone
DeepReinforce shipped the whole family to Hugging Face in GGUF and MLX; vendor scores claim Opus-4.8 parity and the model sets its own harder tasks.
Copy markdown
397B flagship down to a phone-sized 9B
Three sizes ship under MIT: a 397B MoE flagship, a 35B model with 3B active, and a dense 9B with an iPhone/Android build. Weights are on Hugging Face now in BF16, FP8, NVFP4, GGUF and MLX, so the small ones run locally today.
Opus-4.8 numbers, with a big asterisk
The 397B claims 86.0 on SWE-bench Verified, 86.1 on Terminal-Bench 2.1 and 92.8 on GPQA Diamond, matching or beating Claude Opus 4.8 on several tests but trailing on hard agentic work (13.5 vs 21.1 on Frontier-Bench). Every number is vendor-run and none has been reproduced independently yet.
It writes its own harder homework
DeepReinforce credits a self-improvement loop, in which the model proposes progressively harder tasks for itself during training, for the jump over Ornith-1.0. It's the most interesting claim in the release and the one most worth waiting on outside verification for.
262K context, ~1M with YaRN
The native window is 262,144 tokens, extendable to roughly 1M with YaRN scaling at factor 4.0. That's room to drop a mid-sized repo into the flagship, or a long document set into the 9B running on-device.