Ornith-1.5 open-source model family released

Summary

Ornith released its 1.5 open-source LLM family on August 19, 2026. The family consists of a 397B MoE flagship, a 35B-A3B MoE mid-tier model, and a 9B dense model, including a quantized variant for smartphones. All models are available on Hugging Face under the MIT License.

Why it matters

The three-tier open-weight product lineup claims performance on par with Claude Opus 4.8 at the high end. In the mid and small segments, it significantly outperforms larger dense models like Gemma-4-31B and Muse-Glimmer-30B. This situation significantly lowers the barrier for local and on-device deployment. The MIT License and same-day release of quantized versions enable immediate use in commercial and mobile workloads without API dependency.

Highlights

• The 397B MoE flagship claims to perform on par with or better than Claude Opus 4.8 in multiple benchmarks
• The 9B dense model outperforms the much larger Gemma-4-31B in most tests; the quantized 'Mobile' variant targets smartphones
• Training utilizes a self-improvement loop where the model suggests increasingly difficult tasks to itself

Critical take

All performance claims are based on Ornith's own blog post and benchmark screenshots. No independent third-party evaluation is cited. The 'self-improving' training loop is described only with vague expressions. Therefore, it is difficult to assess whether the gains are reproducible or an artifact of benchmark-specific tuning.