Introduction of Ornith-1.0: Open-Source Self-Improving Models for Coding

Ornith-1.0 is a self-improving open-source model for agentic coding.

The model is evaluated against size-appropriate baselines using consistent harnesses and decoding setups across different benchmarks, including Terminal-Bench 2.1 and SWE-bench.

It includes a dense 9B model and two Mixture-of-Experts models (35B and 397B), all of which support a 256K context window. The dense model can run on a single 80GB GPU, while the MoE models require multi-GPU setups.

Ornith-1.0 features a reasoning model that outputs a reasoning block before the final answer, enhancing its ability to perform complex tasks.

The model can be accessed through various interfaces, including OpenAI-compatible APIs, and is optimized for terminal-based coding agents, making it suitable for automating coding tasks and understanding large codebases.