damfle/ornith-1.5-35b-a3b-custom

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 20, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

Ornith-1.5-35B-A3B is a 35.1 billion parameter Mixture-of-Experts (MoE) model developed by the Ornith Team, activating approximately 3 billion parameters per token. This model is a major step towards foundation models through end-to-end self-improvement, continuously generating new training tasks and optimizing solutions via reinforcement learning. It significantly outperforms similar-sized and dense models on coding and agentic benchmarks, making it highly optimized for complex coding tasks and agentic workflows. The model supports a 32,768 token context length, extendable to 1 million tokens with YaRN scaling.

Loading preview...

Ornith-1.5-35B-A3B: Self-Improving Agentic Coding Model

Ornith-1.5-35B-A3B is a 35.1 billion parameter Mixture-of-Experts (MoE) model from the Ornith Team, built upon Qwen3.5 and Gemma4. It represents a significant advancement in self-improvement, moving beyond fixed tasks to continuously generate and optimize solutions through reinforcement learning, encompassing task generation, scaffold construction, and solution rollouts.

Key Capabilities & Performance

  • Efficient MoE Architecture: Activates only ~3 billion parameters per token, offering strong performance with reduced computational overhead compared to dense models.
  • Superior Coding & Agentic Benchmarks: Outperforms models like Qwen 3.6-35B, Gemma 4-31B, and Muse Glimmer-30B across various coding and agentic benchmarks, including Terminal-Bench 2.1, SWE-bench, DeepSWE, and MCP-Atlas.
  • Reasoning Model: Features a built-in reasoning parser that separates chain-of-thought (<think>) from the final answer, enhancing transparency and debugging for complex tasks.
  • Tool Calling: Emits well-formed function calls, parsed into standard tool_calls for seamless integration with agent frameworks.
  • Extended Context Window: Supports a native context window of 32,768 tokens, extendable up to 1 million tokens using YaRN scaling for long-context applications.

Ideal Use Cases

  • Agentic Coding: Excels in scenarios requiring autonomous code generation, debugging, and problem-solving within terminal environments.
  • Complex Software Engineering Tasks: Highly effective for tasks like bug fixing (SWE-bench), code understanding, and repository navigation.
  • Tool-Integrated Workflows: Designed for applications leveraging external tools and APIs, with robust tool-calling capabilities.
  • Research in Self-Improvement: A valuable model for exploring and developing advanced self-improving AI systems.

Finetunes
1 models