Ornith-1.5 Open Weights Score 86.1 on Terminal-Bench

The Ornith team published Ornith-1.5 on August 19, 2026, and the interesting part is not only the benchmark line. Ornith-1.0 taught models to build their own scaffolding, the tool structure wrapped around a task. Ornith-1.5 closes the loop further: according to the team's own announcement, the system now generates its own tasks, builds the scaffolds for them, and produces the solution rollouts it then learns from through reinforcement learning. Self-scaffolding became self-improvement, and the weights are MIT-licensed.

• Ornith-1.5 comes in three sizes: a 397B mixture-of-experts flagship, a 35B MoE, and a 9B dense model
• The 397B model scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, which the team compares to Claude Opus 4.8 at 85.0 and 59.0
• Ornith-1.0's 397B variant posted 77.5 on Terminal-Bench 2.1 in June, so the jump is roughly nine points in under two months
• All weights are MIT-licensed and published on Hugging Face, with a quantized build targeting iOS and Android

What Changed Between Ornith-1.0 and Ornith-1.5?

Ornith-1.0 arrived in late June 2026 from DeepReinforce with a genuinely unusual training idea. Most agentic coding models are trained inside a scaffold that humans designed: a fixed loop of read file, run command, check output. Ornith learned to construct that loop itself, treating the harness as something to optimize rather than something to inherit. The name comes from the Greek for bird, as in a bird building its own nest.

Version 1.5 extends that idea backward into the data pipeline. Rather than only learning how to wrap a given task, the system proposes tasks, scaffolds them, attempts them, and feeds the results into reinforcement learning. That is a tighter loop with fewer human-authored stages in it, and it is the mechanism the team credits for the improvement on terminal-style benchmarks, where a model has to plan several shell operations ahead rather than emit one correct function.

How Good Are the Ornith-1.5 Benchmark Scores?

On Terminal-Bench 2.1, which measures whether a model can complete real command-line engineering tasks end to end, the 397B model reports 86.1. On DeepSWE it reports 56.0. The team's own comparison places that near Claude Opus 4.8 at 85.0 and 59.0 respectively, which is a meaningful claim for an openly licensed model, and one worth confirming against independent runs before anyone rebuilds a pipeline on it. Self-reported benchmarks from a model's authors are a starting point, not a verdict.

The smaller models are arguably the better story for most readers. The 35B mixture-of-experts variant activates a fraction of its parameters per token, and the team reports it beating similarly sized competitors on coding and agentic work. The 9B dense model is put in the neighborhood of Gemma 4-31B. If those hold up, they land in the size class that actually fits on a workstation, which is where open weights change what individual developers can do rather than only what labs can do.

Why MIT-Licensed Weights Matter for Agentic Coding

Licensing is the quiet variable that decides whether an open-weight release is usable at work. MIT is about as permissive as it gets: modify, redistribute, and ship commercially without a separate agreement. For teams that want an agent reading their private repository, the ability to run the model on their own hardware is often the requirement that rules everything else out.

This continues a run of capable open releases we have covered, including Qwen3.8-27B and its 262K-context vision model and the 500K-context agentic coding push behind Grok 4.6. The pattern across all three is the same: the gap between what you can download and what you can only rent keeps narrowing on agentic tasks specifically.

What to Watch Next

Two things will tell us whether 1.5 holds up. The first is independent evaluation on Terminal-Bench and SWE-Bench Verified by people who did not train the model. The second is quantization quality, since a 397B MoE only matters to most builders once it survives compression to something a small cluster can serve. The mobile build the team mentions suggests they are already working that angle.

For more on the local-first side of this shift, see our artificial intelligence coverage.

Sources: Ornith — August 19, 2026; Hugging Face — August 19, 2026; MarkTechPost — June 25, 2026.