GitHub radar
A code-focused AI model trimmed down to run on your own computer
This is a pruned, compressed version of a large Chinese model built for code and tool-calling agents. You get a model that fits on one powerful machine instead of the cloud — at the cost of some capability.
An AI model with half its 'expert' blocks removed and its remaining weights compressed so it fits on one powerful computer.
When it helps
Useful if you want to try a large code model offline on strong hardware. Not for daily serious coding work — it's weaker than cloud assistants.
Pros
- Model file shrinks from 138 GB to 72 GB
- Keeps 94.8% of SWE-bench, 98.7% of LiveCodeBench scores
- Can also handle images via a separate file
- Runs on one computer, no cloud needed
Cons
- Needs 96–128 GB memory and about 200 GB disk space
- Weaker in real agentic work than benchmarks suggest
- No replacement for a cloud coding assistant day-to-day
How to set it up — step by step
- 1Check your computer has 96 GB+ memory (128 GB better) and 200 GB free disk space.
- 2Open a coding AI agent (Claude Code, Codex) in your project folder.
- 3Send the agent the text from the block below.
- 4Let the agent download the model from Hugging Face (a site that hosts AI models) and run it through llama.cpp, a free program for running models.
- 5Test how the model handles your code and tool calls before trusting it with serious work.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the DeepSeek-V4.1-Flash-1017 GSQ-RCO Coder model for me: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF — run it through llama.cpp on my machine and show me how to use it for code and tool calls.
You'll need a machine with 96 GB of RAM or more (128 GB is safer) and about 200 GB of free disk space — a Mac Studio with 128 GB+ of unified memory or a workstation with a large GPU will do. A regular 16-32 GB laptop won't handle it.
Open on Hugging Face▌ More finds