Ornith: Local Agentic Coding Models
AICodeKinggo watch the original →
the gist
Ornith is a new family of open-weights coding models that focus on self-scaffolding for agentic workflows, offering competitive performance in 9B and 35B sizes for local execution.
Self-Scaffolding Agentic Architecture
Ornith distinguishes itself by training models to handle their own agentic scaffolding rather than relying solely on external harnesses. The models are post-trained on Gemma and Qwen architectures to manage planning, error recovery, file manipulation, and tool invocation internally. This approach aims to reduce the common failure modes found in coding agents where the model is strong but the execution loop is poorly integrated.
Local Deployment and Performance
While the 397B parameter model targets high-end benchmarks, the 9B and 35B variants are optimized for local hardware. The author notes that these models are less prone to the output glitches observed in base Qwen 2.5 models. To ensure proper functionality, users must utilize runtimes that correctly parse reasoning tags and tool-call blocks. The author recommends using GGUF versions via LM Studio or Ollama for the easiest setup, noting that the 35B model performs reliably when paired with agent frameworks like OpenCodeInterpreter or Hermes.
Benchmark Context
DeepReinforce reports strong performance on coding benchmarks, with the 397B model scoring 77 on TerminalBench and 82 on SWE-bench. The 35B model achieves 64 on TerminalBench and 75 on SWE-bench, while the 9B model reaches 43 and 69 respectively. The developer claims to mitigate reward hacking by using a fixed environment and an LLM judge to verify that the model is solving tasks rather than gaming the test suite.