By Roman Aglaganov
A team publishing as Ornith AI has released Ornith-1.5-122B on Hugging Face, with compressed versions for smaller hardware posted on 20 October. It is the newest member of a family that also includes 9-billion, 35-billion and 397-billion-parameter models, released in August.
What it is
The model is a mixture of experts, which uses only part of its 122 billion parameters for each word, and its makers say it is designed to run on a single GPU. It is released under the MIT licence, which lets anyone use, change and sell products built on it. It is a reasoning model: it thinks through a problem before answering.
The training idea
Ornith-1.0 was built on top of open models from Alibaba (Qwen3.5) and Google (Gemma 4). For 1.5, the team says it extended a “self-improvement loop”: rather than relying on a fixed set of tasks written by people, the model generates new training tasks, builds the tools to tackle them, tries solutions and learns from the results through reinforcement learning.
The claims
On its model card, the team reports 84.5% on SWE-bench Verified and 76.9% on SWE-bench Pro, tests of fixing real software bugs, with safeguards meant to stop the model gaming the tests. It does not lead everywhere: on TerminalBench 2.1 it scores 81.2, below Qwen3.8-Flash-Next and DeepSeek-V4-Flash in the same table. These are the developers’ own results.
Why it matters
Open models that companies can run on their own hardware keep getting more capable, and a model that can invent its own practice problems needs less expensive human-written data. If that approach holds up under outside testing, it lowers one of the biggest costs of building a strong model.
