Ornith AI released Ornith-1.5-397B. It is a 397B mixture-of-experts reasoning model. It extends Ornith-1.0 by jointly optimizing task generation, scaffold construction and solution rollouts. It scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE. It matches Claude Opus 4.8 on coding benchmarks. It outperforms GLM-5.2 and DeepSeek-V4-Flash-0731. It supports a 262,144 token context window. It can be extended to ~1M tokens with YaRN RoPE scaling. It emits OpenAI-compatible tool calls and reasoning traces. It works with vLLM, SGLang, Ollama, Atomic.chat, llama.cpp and OpenCode.
What changed
Ornith-1.0 used fixed human-curated tasks and manual harnesses for training.
What it unlocks
Running a 397B mixture-of-experts model locally with end-to-end self-improvement for coding and reasoning tasks.
- 397B parameters
- 262,144 token context window
- 86.1 on Terminal-Bench 2.1
- 56.0 on DeepSWE
What you need to act on it
- multi-GPU setup or quantized build
- Transformers ≥ 5.8.1, vLLM ≥ 0.19.1, or SGLang ≥ 0.5.9
Sources