Published: 20 August 2026
Ornith released Ornith-1.5 on 19 August 2026, an open-weight model family under the MIT licence in 397B and 35B mixture-of-experts and 9B dense sizes. On Ornith’s own evaluations the 397B scores 86.1 on Terminal-Bench 2.1 against 85.0 for Claude Opus 4.8. Weights for every size are already downloadable from Hugging Face.
What is Ornith-1.5 and how was it trained?
Ornith-1.5 is a coding and agentic model family from DeepReinforce, the team that shipped Ornith-1.0 in June 2026. It comes in three sizes: a 397B mixture-of-experts flagship, a 35B mixture-of-experts model activating 3B parameters per token, and a 9B dense model with a quantised Mobile build for iPhone and Android.
The training method is what Ornith is selling. Instead of a fixed set of human-written tasks, the model proposes its own, builds a scaffold for each, and generates the solution rollouts that become the reinforcement learning signal. Task reward multiplies validity, novelty and difficulty, with difficulty targeting a 0.2 success rate.
Ornith-1.5 benchmarks, context window and licence
- The 397B posts 86.0 on SWE-bench Verified,
- 92.8 on GPQA Diamond,
- 86.6 on BrowseComp.
Its context window is 262,144 tokens, extendable to roughly 1M with YaRN scaling at factor 4.0. Every figure comes from Ornith’s own runs, averaged over five attempts, with the harness and sampling settings published per benchmark.
- The 35B reaches 68.5 on Terminal-Bench 2.1,
- 79.0 on SWE-bench Verified on 3B active parameters.
- The 9B reaches 47.0 on Terminal-Bench 2.1,
- 70.6 on SWE-bench Verified.
All sizes ship under the MIT licence in BF16, FP8, NVFP4, GGUF and MLX builds. There is no hosted API and no published price list.
How does Ornith-1.5 compare to Claude Opus 4.8 and Kimi K3?
Against Claude Opus 4.8 the result is mixed rather than a clean win. Ornith-1.5-397B leads on Terminal-Bench 2.1, SWE-bench Verified, WideSearch and SWE-bench Multilingual, and trails on the harder agentic rows: 13.5 against 21.1 on Frontier-Bench v0.1, 59.5 against 69.7 on NL2Repo, and 56.0 against 59.0 on DeepSWE.
Kimi K3 still leads the open-weight field at 88.3 on Terminal-Bench 2.1, but it is a 2.8T model against 397B. Against GLM-5.2 at 81.0 and DeepSeek-V4-Flash-0731 at 82.7, Ornith-1.5 wins at similar or smaller scale. These are all vendor numbers and none has been independently reproduced yet.
Who built Ornith-1.5 and how do you run it?
Ornith is the model line from DeepReinforce, which shipped Ornith-1.0 in June 2026 under MIT across 9B, 31B, 35B and 397B variants, post-trained on Gemma 4 and Qwen 3.5 checkpoints. Reward hacking has been a running theme in the team’s published reinforcement learning work, and Ornith-1.5 carries that defence into task generation.
Serving needs recent runtimes: Transformers 5.8.1, vLLM 0.19.1 or SGLang 0.5.9 and above. The published vLLM recipe for the 397B assumes eight-way tensor parallelism on one node, for example eight H200 141GB cards. GGUF builds run through Ollama and llama.cpp, and the model exposes OpenAI-compatible tool calling for agent CLIs such as OpenCode.
What this means
This is worth testing now if you self-host a coding agent. An MIT licence on a model that matches a closed flagship on Terminal-Bench removes both the per-token bill and the usage terms. The 35B is the more interesting build: near-flagship agentic coding on hardware a mid-sized engineering team can actually afford to serve.
Two cautions. The 397B needs roughly 800GB in BF16 and eight-way tensor parallelism, so it is a cluster commitment rather than a weekend experiment. And a model trained on tasks it wrote itself has an obvious failure mode, which is why the gap on Frontier-Bench and NL2Repo matters more than the headline score.
Ornith’s announcement, the full benchmark tables and the model cards are published at the official announcement of Ornith 1.5 on the Ornith blog.
