Ornith-1.5 Beats Claude Opus on Coding Benchmarks and Runs Locally

Ornith-1.5-397B is a 397-billion-parameter mixture-of-experts model published on Hugging Face that reaches Claude Opus territory on agentic coding benchmarks, built around a training method where the model writes its own curriculum. The GGUF release has crossed 897k downloads, and quantized builds run locally on llama.cpp, Ollama, LM Studio, and vLLM.

On Terminal-Bench 2.1 it scores 86.1 against Claude Opus 4.8's 85.0, and 56.0 on DeepSWE against Claude's 59.0, while outperforming open-source models of similar scale including GLM-5.2 (82.7 and 46.2) and DeepSeek-V4-Flash-0731 (82.7 and 54.4). The jump from Ornith-1.0 to 1.5 on DeepSWE is particularly sharp, moving from 8 to 56, which suggests the self-generated curriculum found gaps the earlier version could not close at all.

How the self-improvement loop works

Ornith-1.5 extends the prior version by expanding a self-improvement loop from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts. Rather than a fixed set of human-curated tasks and manually designed harnesses, the system continuously generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning.

Each training cycle runs in three stages. Given an environment or codebase, high-level instructions about the task type, and access to the model's previous task-solving history, the system proposes progressively harder tasks that expose capability gaps. The model then generates or refines a task-specific scaffold: the instructions, tools, decomposition strategy, and orchestration used to approach the problem. Reward flows back through all three stages, so the model improves at inventing tasks, not just solving them.