Ornith-1.5
#25 in Open-Source LLMsOrnith · v1.5 · 5× · last seen Aug 21, 2026
Ornith-1.5 is an Open-Weight language model family released on August 19, 2026 by Ornith AI (DeepReinforce) under the MIT License, consisting of a 9B Dense model, a 35B MoE model (approximately 3B active parameters per Token), and a 397B MoE flagship. The models are trained through an "end-to-end self-improving" training procedure in which the model generates its own training tasks, scaffolds, and solution rollouts and learns from them via Reinforcement Learning. All sizes support a context window of 262,144 Token (expandable to approximately 1M via YaRN) and are available in several quantized formats (FP8, GGUF, MLX, NVFP4) on Hugging Face. There is no official hosted API offering and no published price list, as it is a pure Open-Weight model.
Features
| Key Benchmark (%) | 397B flagship: 86.1% on Terminal-Bench 2.1 (vs. 85.0% Claude Opus 4.8) |
| Context Window (Tokens) | 262,144 tokens, extendable to roughly 1M tokens via YaRN scaling (factor 4.0) |
| License | MIT license (all sizes: 9B, 35B MoE, 397B MoE) |
| Multimodality | Primarily text-based; Ollama catalog lists image input (vision) support for the 9B model |
| Platform | Open-weight models on Hugging Face; runs via vLLM, SGLang, Ollama, LM Studio; agent integration via OpenCode and Hermes Agent; fine-tuning via Unsloth |
| Price per 1M Tokens | No hosted API/no price list – pure open-weight release for self-hosting |
| Release Date | August 19, 2026 |