Overview
Nex-N2-mini is a 35.1B-parameter open-weight model from nex-agi, built on the Qwen3.5-35B-A3B-Base.
- License: Apache 2.0
- Context length: 256K tokens
- Languages: Multilingual
- Minimum hardware: ~20 GB VRAM
- Strengths: reasoning and on-device inference
Highlights
- Agentic coding — 74.4 on SWE-Bench Verified and 60.7 on Terminal-Bench 2.1.
- Hard reasoning — 82.6 on GPQA Diamond with explicit reasoning traces.
- Vision + long context — image-text input and a 256K-token window.
- A3B MoE — ~3B active parameters keep memory modest for a 35B model.
What it is good at
Nex-N2-mini carries the tools, thinking, vision, reasoning, and code capabilities of the Nex-N2 line. Its strongest results are in agentic coding and tool use.
- Agentic coding — it scores 74.4 on SWE-Bench Verified and 60.7 on Terminal-Bench 2.1, so it can read a repo, edit files, and run terminal commands across multi-step tasks.
- Tool calling and reasoning — function calling plus explicit reasoning traces let it plan, call a tool, read the result, and adjust. It reaches 82.6 on GPQA Diamond for hard reasoning questions.
- Vision and long context — the model accepts image-text-to-text input and handles a 256K-token context, useful for screenshots, diagrams, and large codebases or documents in one session.
Running it locally
The model is 35.1B parameters total with a 256K context window, but the 3B active-parameter MoE keeps memory modest. At a Q4 quantization the weights sit around 21 GB, which fits a 24 GB GPU like an RTX 3090, 4090, or 5090, or a Mac with 32 GB or more of unified memory. With CPU offloading through llama.cpp it has been run on far smaller cards at lower speeds.
huggingface-cli download nex-agi/Nex-N2-mini
You can serve the full-precision weights with Transformers or vLLM, or use a GGUF quant in a llama.cpp-based runtime. In Atomic Chat the model loads with one click, with no server setup or command line needed.
License
Nex-N2-mini is released under the Apache-2.0 license. That permits commercial use, modification, redistribution, and private deployment, as long as you keep the license and attribution notices. You can fine-tune it and ship it inside your own products without a separate agreement.