DeepReinforce released Ornith-1.5 on August 19, 2026, an open-weight family under MIT license spanning three sizes: a 397B mixture-of-experts flagship, a 35B MoE activating only 3B parameters per token, and a 9B dense model with a mobile-quantized variant. The headline feature is a closed self-improvement training loop where the model does not just solve human-authored tasks—it proposes its own tasks, builds task-specific scaffolds to tackle them, and generates reinforcement learning rollouts that feed back into training.
On benchmarks, the 397B flagship scores 86.1 on Terminal-Bench 2.1 and 56 on DeepSWE, matching or slightly exceeding Claude Opus 4.8 at 85.0 and 59.0. The 35B MoE outperforms larger dense models like Gemma 4-31B and Meta's Muse Glimmer on agentic coding (68.5 vs. 43.4 on Terminal-Bench 2.1). The 9B model runs on a single GPU with quantized builds targeting iPhone and Android. All weights are available on Hugging Face with day-one support in GGUF, MLX, FP8, and NVFP4 formats.
For builders of coding agents and agentic AI, Ornith-1.5's significance is not just the benchmark parity with Claude Opus. It is that self-improvement loop—where task generation, scaffolding, and solution rollouts all train together—removes the fixed human-curated task set. Independent benchmark verification is still pending, but the model has already hit Hacker News and shipped with production support in vLLM, Ollama, and OpenCode. The 35B is where you deploy.