Ornith 1.5 brings self-improving open AI to local devices
August 20, 2026
Ornith has released three open models ranging from 9 to 397 billion parameters. Most relevant: the smallest model targets local GPUs and, in quantized form, mobile devices.
What this is about
The research group Ornith released the Ornith 1.5 model family on August 19, 2026. It includes a dense 9B model and two mixture-of-experts models with 35 and 397 billion total parameters. Weights and variants are available on Hugging Face, and the official model cards specify the MIT license.
The notable part is not another leaderboard but the range: Ornith offers a large research model, a 35B model that activates only about three billion parameters per token, and a 9B model for a single GPU. Quantized versions of the 9B model are also intended to reach mobile devices, according to the developer.
What Ornith 1.5 actually does
Ornith 1.5 builds on Ornith 1.0 and expands its training loop. The model does not merely solve a fixed set of tasks. During training, the system generates new tasks, searches for suitable solution strategies, and evaluates the resulting rollouts. Those three parts jointly feed further reinforcement learning.
The family covers three size classes. Ornith 1.5 397B is a large MoE model. Ornith 1.5 35B-A3B has 35 billion total parameters but activates about three billion per token. Ornith 1.5 9B is a dense model for local use. Official cards list a 256,000-token context for several variants and provide Transformers, GGUF, MLX, FP8, and NVFP4 formats.
Why it matters
Open models become practical when developers do not need an external API for every task. Local execution can keep confidential code on the user's machine, make costs more predictable, and enable experiments without dependence on one provider. A 9B model is considerably more accessible than a model with hundreds of billions of parameters.
Ornith reports a score of 79 on SWE-bench Verified and 68.5 on Terminal-Bench 2.1 with Claude Code for the 35B model. For the 9B model, the group reports 70.6 and 47 respectively. These are vendor-reported results from the model cards, not independent reproductions. They show the target of the release: coding, tool use, and agent workflows.
In plain language
Instead of giving a cook the same recipes repeatedly, Ornith lets the cook propose new dishes during training, develop suitable kitchen routines, and judge the outcomes. The goal is to improve not only the recipe collection but also how difficult work is prepared. An independent kitchen still needs to test whether guests actually like the food.
A practical example
A software team wants to triage 40 internal bug reports each day and create repair suggestions for five of them. It runs Ornith 1.5 9B on a local GPU, limits access to a test repository, and checks every patch through tests and human review. Source code and tickets therefore remain inside the company network.
After two weeks, the team compares 200 suggestions with its previous model: How many pass the tests, how often does the model edit the wrong files, and how much compute time does a useful patch require? Only this internal measurement can show whether published benchmark results matter in the real project.
Scope and limits
- Most performance figures come from Ornith. Independent reproductions were not yet available on August 20, 2026.
- An MIT license and open weights do not make training data, safety behavior, or every risk fully transparent.
- A locally running model is not automatically safe. Coding agents still need sandboxes, least privilege, tests, and human approval.
- The 397B variant remains impractical for most individuals. Even quantized mobile variants may be slower, less accurate, or more power-hungry than server models.
SEO & GEO keywords
Ornith 1.5, Ornith AI, open-weight model, local AI, coding model, reinforcement learning, self-improving AI, Hugging Face, MIT license, GGUF, MLX, SWE-bench
💡 In plain English
Ornith 1.5 is an open model family for coding and tool use. Its smallest variant is designed to run locally on one GPU and, when quantized, even on mobile devices, but its advertised performance still needs independent verification.
Key Takeaways
- →Ornith released the model family on August 19, 2026 under the MIT license.
- →The family includes 9B, 35B-A3B, and 397B variants.
- →The 9B model targets local single-GPU use and quantized mobile variants.
- →Its training loop jointly generates tasks, solution strategies, and rollouts.
- →The benchmark figures are vendor-reported and not yet independently confirmed.
FAQ
Is Ornith 1.5 open source?
The weights and model cards are publicly available and specify the MIT license. That does not automatically provide full transparency about training data and the training process.
Can Ornith 1.5 run locally?
The 9B model targets a single GPU. Additional GGUF and MLX variants make local and quantized use easier.
Are the benchmarks independently confirmed?
Not yet. The cited numbers come from official model cards and should be treated as vendor-reported results.
What does A3B mean for the 35B model?
The model has 35 billion total parameters but activates about three billion per token. This is intended to reduce compute compared with a fully active 35B model.