A model can improve not only by solving more examples, but by creating better examples and better ways to solve them. That is the key shift here: Ornith-1.5 trains task generation, scaffold construction, and solution rollouts together, turning the process that produces its learning experiences into part of the learning loop itself.
Key Capabilities
- Joint self-improvement: The model generates tasks, builds task-specific instructions and tool workflows, then produces solutions that are scored through reinforcement learning. This reduces reliance on a fixed set of curated tasks and manually designed harnesses.
- Agent-oriented performance: It is designed for coding and tool-using work, with reported results across coding, reasoning, and agent benchmarks. The model card reports 84.5 on SWE-bench Verified and 79.4 on ClawEval; these are benchmark results, not guarantees of performance on every codebase or workflow.
- Large model, constrained footprint: Its 122B mixture-of-experts design is presented as deployable on a single 80GB GPU, with a 262,144-token context window. A quantized mobile variant is also listed for edge deployment.
Who It Suits
A good fit for teams evaluating open-weight models for coding agents, long-context repository work, or tool-using workflows who can operate modern inference runtimes. Look elsewhere if you need a small, low-resource model without quantization, or if your evaluation depends on capabilities not covered by the published benchmarks. The reported scores come from specific harnesses and settings, so test with your own tasks before choosing it.
