Skip to content
Neural Daily

ModelsNews

Ornith-1.5-122B: open coding model for a single GPU

Ornith has released Ornith-1.5-122B, an MIT-licensed 122-billion-parameter MoE reasoning model. Its model card says it serves on a single 80GB GPU and scores 84.5 on SWE-bench Verified.

By: Mikola Máté Patrik · · 2 min read

Ornith 1.5 122B performance evaluation bar charts comparing coding, reasoning and agentic benchmarks
Image: Ornith · Press kit · Official model card benchmark chart · Source

What shipped

Ornith-1.5-122B succeeds Ornith-1.0, which the company says was built on Qwen3.5 and Gemma4 with continued pretraining, mid-training and post-training (model card). The new version extends its self-improvement loop to jointly optimize task generation, scaffold construction and solution rollouts. Weights are MIT-licensed and the context window is 262,144 tokens.

Numbers from the model card

  • SWE-bench Verified: 84.5 (Qwen3.8-Flash-Next 125B: 82.8).
  • SWE-bench Pro: 76.9 (Qwen3.8-Flash-Next: 43.9).
  • TerminalBench 2.1: 81.2, where Qwen3.8-Flash-Next (83.4) and DeepSeek-V4-Flash (82.7) score higher.
  • BrowseComp: 78.5, MCP-Atlas: 74.6, HLE with tools: 48.0.
Ornith 1.5 122B performance evaluation bar charts comparing coding, reasoning and agentic benchmarks
Ornith · Press kit · Official model card benchmark chart · Source

Source: Ornith, Hugging Face model card

These are the developer's own measurements with no independent check yet. The card documents harness settings and says anti-reward-hacking safeguards were applied to the SWE-bench runs.

How to run it

An OpenAI-compatible server starts with vLLM (0.19.1 or later) or SGLang (0.5.9 or later), with tool calling and reasoning parsers wired in. FP8, MLX and GGUF builds are available for Macs and local machines. The model opens with a <think> block by default, and the reasoning comes back in a separate reasoning_content field.

For teams that want a coding agent on their own infrastructure, this is one of the strongest open options at this size. Test it on your own repository before swapping out an existing model.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles