GitHub radar
Open agentic model for browser and terminal control
A 35B open-weights model trained to surf the web, use terminal, run & test code, reason about images, and invoke external utilities.
This model (by Nex-AGI) is a ‘next generation family of agentic models designed for long horizon tasks in the real world’. It can browse the internet, interact with the command line, execute & test code, reason about images, and call external tools. There are three reasoning levels (quick, average, thorough) depending on the complexity of the task. The model scores 71.2 on OSWorld-Verified and 43.8 on SWE-Bench Pro. A quantized version of the model is available for llama.cpp, Ollama, and LM Studio
Why a vibe-coder should care
Want to mess around with an open source computer use agent? Then this is one of the best open weight models out there right now. Be warned though, this model is quite large (even the quantized version is ~21GB). You’ll need either a Mac with 64GB Unified Memory or a cloud GPU instance to use it (it won’t work on a normal laptop).
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Find a quantized version of Nex-AGI's Nex-N2.5-mini that fits my RAM and install it via Ollama or LM Studio: https://huggingface.co/nex-agi/Nex-N2.5-mini — run it and show me how to use it
The 4-bit quantized version requires around 21 GB — a Mac with 64 GB unified memory or a rented cloud GPU. The authors officially recommend a dual H100 setup.
Open on Hugging Face▌ More finds