Ornith-1.0 Release: Self-Improving Open-Source Models for Agentic Coding

Ornith-1.0 is a series of open-source models designed for agentic coding, utilizing a self-improving training framework to optimize both solution rollouts and the scaffolding that drives them. These models are post-trained on top of Gemma 4 and Qwen 3.5 and are available under the MIT license.

Model Architecture and Availability

Ornith-1.0 is released in several sizes and architectures to balance performance and hardware requirements. All models support a 256K token context window and provide an OpenAI-compatible interface.

  • 9B-Dense: Optimized for single-GPU serving and fine-tuning. Fits on a single 80GB GPU.
  • 35B-MoE: A Mixture-of-Experts model designed for multi-GPU nodes.
  • 397B-MoE: The largest variant, designed for high-performance serving on multi-GPU nodes.

Checkpoints are available in bf16 and FP8 (for reduced VRAM usage), as well as GGUF formats for local inference via llama.cpp or Ollama.

Self-Improving Training Framework

Ornith-1.0 employs Reinforcement Learning (RL) to achieve self-improvement. Unlike traditional training, this framework jointly optimizes the scaffold (the process/strategy used to find a solution) and the resulting solution. By learning to generate better search trajectories, the model discovers higher-quality solutions for complex coding tasks.

Performance Benchmarks

Ornith-1.0 demonstrates state-of-the-art performance among open-source models of comparable size across several agentic coding benchmarks.

Ornith-1.0-397B Highlights

Compared to other large-scale models, the 397B variant shows strong results in terminal-based coding:

  • Terminal-Bench 2.1 (Terminus-2): 77.5%
  • SWE-bench Verified: 82.4%
  • SWE-bench Pro: 62.2%
  • NL2Repo: 48.2%

Ornith-1.0-35B Highlights

  • Terminal-Bench 2.1 (Terminus-2): 64.2%
  • SWE-bench Verified: 75.6%
  • Claw-eval Avg: 69.8%

Ornith-1.0-9B Highlights

  • Terminal-Bench 2.1 (Terminus-2): 43.1%
  • SWE-bench Verified: 69.4%
  • NL2Repo: 27.2%

Deployment and Integration

Ornith-1.0 is a reasoning model that outputs a <think> block before the final answer. It is compatible with major serving runtimes including vLLM (>= 0.19.1), SGLang (>= 0.5.9), and Transformers (>= 5.8.1).

Agent Framework Compatibility

Because it exposes an OpenAI-compatible endpoint with tool-calling capabilities, Ornith-1.0 integrates with:

  • OpenHands: Via LiteLLM using the openai/ prefix.
  • Hermes Agent: By pointing to the local Ornith server.
  • OpenClaw: Via OpenAI-compatible endpoint configuration.
  • Coding CLIs: Optimized for terminal-based agents like OpenCode.

Community Insights and Technical Feedback

User feedback from the developer community highlights both the strengths and weaknesses of the model family:

Performance and Efficiency

Some users report that the 35B model is faster and more efficient than Qwen 3.6 35B, producing shorter chains of thought while maintaining high utility for medium-sized codebases.

"From what I personally tested Ornith-1.0 35B is slightly better than Qwen-3.6 35B... the model is way faster than Qwen3.6 35B. It seems Ornith produce a smaller chain of thought."

Limitations and Criticisms

Other users have expressed skepticism regarding the "self-improving" claims, suggesting the models are "benchmaxxed" versions of Qwen or Gemma. Some reported issues with hallucinations in chat mode without tools and failures in very long-session tool calls.

"Poor performer here, only found the one bug that almost every model found, despite its performance on other benchmarks being excellent for its size. […] It also performs poorly in a chat without tools, exhibiting an enthusiasm for hallucination."

Additionally, community members noted a discrepancy in the documentation, mentioning a 31B dense model that has not been released or benchmarked.

Sources

Related