Is Ornith-1.5 free to use commercially?

Ornith-1.5 ships under the MIT license on Hugging Face, so the weights can be used, modified and redistributed commercially without a separate agreement. Every size in the family is covered, including the 397B flagship and the quantized FP8, NVFP4, GGUF and MLX builds. There is no hosted API tier or published price list from the Ornith Team.

How much better is Ornith-1.5 than Ornith-1.0?

Ornith-1.5-397B improves on Ornith-1.0-397B across the board in the published table: Terminal-Bench 2.1 goes from 77.5 to 86.1, SWE-bench Verified from 82.4 to 86.0, and Toolathlon-Verified from 43.2 to 71.2. The largest jump is DeepSWE, which moves from 8 to 56.

Can Ornith-1.5 run on one machine?

The Ornith-1.5 family includes a 9B dense model and a 9B-Mobile variant aimed at edge deployment, plus GGUF and MLX builds that target laptops and Apple silicon. The 397B and 35B mixture-of-experts models are server-scale: the published vLLM recipe for the 397B uses eight-way tensor parallelism.

How long a context does Ornith-1.5 handle?

Ornith-1.5-397B has a 262,144-token context window, and the model card notes it can be extended to roughly 1M tokens using YaRN scaling at factor 4.0. The published vLLM and SGLang launch commands both set the maximum length to 262,144 tokens by default.

Where does Ornith-1.5 still lose to Claude Opus 4.8?

Claude Opus 4.8 stays ahead of Ornith-1.5-397B on the harder agentic and reasoning rows of the same table: Frontier-Bench v0.1 (21.1 against 13.5), NL2Repo (69.7 against 59.5), DeepSWE (59.0 against 56.0) and HLE without tools (49.8 against 44.6). Ornith-1.5 leads on WideSearch, ClawEval and SWE-bench Multilingual.

Ornith · 2026-08-19 · major

Ornith-1.5 — open MIT model matches Claude Opus 4.8 on Terminal-Bench

Ornith-1.5 is a family of open-weight models released under MIT in 397B and 35B mixture-of-experts and 9B dense sizes. The 397B scores 86.1 on Terminal-Bench 2.1 against 85.0 for Claude Opus 4.8.

An open-weight model family that writes its own training tasks, then trains on them.

Quick facts

Maker	Ornith Team
Model sizes	397B MoE, 35B MoE, 9B dense
License	MIT
Context window	262,144 tokens
Availability	Open weights on Hugging Face
Formats	BF16, FP8, NVFP4, GGUF, MLX
What's new	Self-improvement loop replaces fixed training tasks

Benchmarks

Terminal-Bench 2.1 (Terminus-2)

Kimi K3		88.3%
Ornith-1.5-397B		86.1%
Claude Opus 4.8		85%
DeepSeek-V4-Flash-0731		82.7%
GLM-5.2		81%
Ornith-1.0-397B		77.5%

source ↗

SWE-bench Verified

Kimi K3		86.2%
Ornith-1.5-397B		86%
Claude Opus 4.8		85.8%
GLM-5.2		83%
Ornith-1.0-397B		82.4%
DeepSeek-V4-Flash-0731		81.6%

source ↗

What is it?

Ornith-1.5 extends the self-scaffolding idea from Ornith-1.0 into a full self-improvement loop, where the model invents its own training work instead of drawing from a fixed human-written set. It comes in three sizes — 397B mixture-of-experts, 35B mixture-of-experts and 9B dense — plus a 9B-Mobile build, all under MIT on Hugging Face.

How does it work?

The loop has three steps that feed each other: the model proposes new tasks, builds a scaffold for each one, and produces solution rollouts that become the reward signal for reinforcement learning. The Ornith Team describes the effect as compounding — 'stronger policies enable the generation of harder and more informative tasks, evolving scaffolds discover better ways to elicit the model's capabilities, and higher-quality rollouts provide increasingly effective learning signals.'

Why does it matter?

Matching a closed flagship on Terminal-Bench while shipping under MIT means a team can run this coding agent on its own hardware with no per-token bill and no usage terms to negotiate. The 9B and 35B builds matter more day to day: they put the same training recipe inside a budget most teams can actually serve.

Who is it for?

teams self-hosting a coding agent

Frequently asked questions

Is Ornith-1.5 free to use commercially?: Ornith-1.5 ships under the MIT license on Hugging Face, so the weights can be used, modified and redistributed commercially without a separate agreement. Every size in the family is covered, including the 397B flagship and the quantized FP8, NVFP4, GGUF and MLX builds. There is no hosted API tier or published price list from the Ornith Team.
How much better is Ornith-1.5 than Ornith-1.0?: Ornith-1.5-397B improves on Ornith-1.0-397B across the board in the published table: Terminal-Bench 2.1 goes from 77.5 to 86.1, SWE-bench Verified from 82.4 to 86.0, and Toolathlon-Verified from 43.2 to 71.2. The largest jump is DeepSWE, which moves from 8 to 56.
Can Ornith-1.5 run on one machine?: The Ornith-1.5 family includes a 9B dense model and a 9B-Mobile variant aimed at edge deployment, plus GGUF and MLX builds that target laptops and Apple silicon. The 397B and 35B mixture-of-experts models are server-scale: the published vLLM recipe for the 397B uses eight-way tensor parallelism.
How long a context does Ornith-1.5 handle?: Ornith-1.5-397B has a 262,144-token context window, and the model card notes it can be extended to roughly 1M tokens using YaRN scaling at factor 4.0. The published vLLM and SGLang launch commands both set the maximum length to 262,144 tokens by default.
Where does Ornith-1.5 still lose to Claude Opus 4.8?: Claude Opus 4.8 stays ahead of Ornith-1.5-397B on the harder agentic and reasoning rows of the same table: Frontier-Bench v0.1 (21.1 against 13.5), NL2Repo (69.7 against 59.5), DeepSWE (59.0 against 56.0) and HLE without tools (49.8 against 44.6). Ornith-1.5 leads on WideSearch, ClawEval and SWE-bench Multilingual.

Try it

vllm serve ornith-ai/Ornith-1.5-397B --tensor-parallel-size 8 --max-model-len 262144