Ornith 1.5: Model Family, Sources and Deployment Questions

Ornith-1.5 is a model family described in the Ornith research announcement. Its opening description lists 397B MoE, 35B MoE and 9B dense scales, including a quantized mobile-oriented 9B version. The announcement links to the project's Hugging Face organization for actual model releases.

What “self-improvement” means here

The article describes a training process in which tasks, scaffolds and solution rollouts are generated and used for reinforcement learning. This is a research/training claim. It does not establish that an ordinary downloaded chatbot silently retrains itself while you use it.

The provider reports coding and reasoning benchmarks. Treat those as results under its stated evaluation configuration, rather than a guarantee of your latency, quality or hardware fit. A model's total parameter count alone does not specify the memory needed by your selected quantization and runtime.

Choose a concrete checkpoint

Before downloading, check:

• The exact repository and revision.
• The model card's base model, licence and intended runtime.
• File size, quantization and the runtime's current support.
• Context length and memory needs for your intended workload.

The 35B-A3B model card appeared in the reviewed search results; verify its current details directly before using it. This page does not infer a licence or memory requirement for the whole family from one search result.

Check the result you need

Choose a small task with a known answer or a result you can inspect. Keep the checkpoint and runtime settings with the output. If comparing models, use the same task and retain failures as well as successes.

Research announcement checked October 30, 2026. Varwo has not downloaded, quantized, served or benchmarked an Ornith model. Availability and hardware suitability require checking the specific release you choose.