GitHub radar
Ornith-1.5: a locally-trained agent with cloud-accuracy
ornith-ai’s Ornith-1.5 agentic coding model — 79% on SWE-bench Verified, equivalent to Claude Opus 4.6 on real codebase tasks, wrapped up as GGUF by peculiar-ragdoll for local use with Ollama.
Ornith-1.5 is an agentic coding model developed by ornith-ai using iterative self-improvement. It was trained by solving real codebase problems, getting feedback, and iteratively improving. Quantized to 4-bits, it’s 22 GB and the fastest coder in its category, equivalent to Claude Opus 4.6 medium on real codebase tasks. It has the highest quality multi-turn conversations of any local model. SWE-bench Verified: 79%, Terminal-Bench: 67–68%, MCP-Atlas (real agentic tasks w/ tools): 70.2%. Context window: 262K tokens (~1M with YaRN). Native tool calls, MCP, and OpenAI API compatible. peculiar-ragdoll has wrapped the original ornith-ai weights in GGUF for local use with Ollama, llama.cpp, etc.
Why a vibe-coder should care
Local coding agents keep your code local, perfect for proprietary or confidential projects. This is no sacrifice, since this model is equivalent to Claude Opus 4.6 on real codebase tasks. It can handle large codebases out of the box (no chunking) and is fully compatible with MCP (drop-in compatible with your existing agentic workflow).
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Download and run Ornith-1.5 by ornith-ai in GGUF format: https://huggingface.co/peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF — pick the right quantized variant for my Mac/GPU, set it up via Ollama, and show me how to use it for coding tasks.
A Mac with 32 GB of unified memory or more, or a GPU with 24 GB — the 4-bit version takes around 22 GB.
Open on Hugging Face▌ More finds