GitHub radar
Nemotron 3.5 Bolt: NVIDIA's local agentic model
This is a compressed version of NVIDIA's large model, built for agentic tasks — working in a browser, terminal, and with tools. You can run it on your own machine and see an agent act on its own.
A model that works on its own in a browser and terminal, calls tools, and processes text together with images — built for autonomous AI agents, not casual chat.
When it helps
Useful if you want to try an agentic model locally on your own hardware and see how it handles tools. It's not optimized for casual conversation — it's built for agentic tasks; in agentic benchmarks it beats comparable open models like Qwen and Gemma.
Pros
- NVIDIA compressed it from 120B to 74B parameters with little quality loss
- Beats Qwen and Gemma on agentic benchmarks (GPQA 86.43 vs 83.40 and 79.61)
- Holds up to 256,000 tokens of context — handles a lot of data at once
Cons
- No comparison with Opus or Fable, the models used daily in Claude Code
- Needs a GPU with 24GB+ memory or a Mac with 32GB+ — a regular laptop won't do
How to set it up — step by step
- 1Check you have a GPU with 24GB+ memory or a Mac with 32GB+ memory — otherwise it won't run
- 2Open your AI agent (Claude Code, Codex) in your project folder
- 3Send the agent the text from the block below
- 4Give it a task in a browser or terminal and watch it work on its own
- 5Check the result and compare it with what you're used to
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install the Nemotron 3.5 Bolt model by NVIDIA through Ollama: ollama run nemotron-3.5-bolt. You need a GPU with 24 GB of memory or more, or a Mac with 32 GB of memory or more. Show me how to give it a task in a browser or a terminal and what it does on its own.
You need a GPU with 24 GB of memory or more, or a Mac with 32 GB of memory or more. NVIDIA's own setup targets professional cards such as the A100 or newer, so this is clearly not the level of a regular laptop.
Open on Hugging Face▌ More finds