Nemotron 3.5 Bolt: Free Local Model & Setup

Nemotron 3.5 Bolt is NVIDIA’s free, open-weight 74B agent model that fits in 24GB of memory. See what it needs, how to run it with Ollama and its limits.

NVIDIA released Nemotron 3.5 Bolt on October 15, 2026, alongside the bigger Nemotron 3.5 Super VL. Bolt is the one most people searching for “nemotron 3.5” care about, because it’s built to run on a single desktop GPU instead of a server. This page explains what Bolt is, what hardware you really need, how to run it today and where it falls short, using NVIDIA’s announcement, the official model card and the Ollama and Unsloth guides, all checked on October 20, 2026.

What is Nemotron 3.5 Bolt?

Nemotron 3.5 Bolt 74B-A9B 24GB is an open vision-language model for local AI agents. It reads text, images and video and answers in text. It has 74 billion parameters in total, but only 9 billion are active for each token, which is why it runs much faster than a dense model of the same size.

The trick that makes it fit on one GPU is a native 2-bit quantization trained into the model (quantization-aware training), so the packed model is about 24GB. NVIDIA built Bolt by compressing Nemotron 3.5 Super VL, and it keeps Super VL’s switchable reasoning mode: you can turn thinking on for harder tasks or off for quick replies.

• Architecture: a Mamba-2 and Transformer hybrid mixture-of-experts (LatentMoE) with a C-RADIO v4-H vision encoder.
• Context: up to 256K tokens on the model card. Usable context depends on your memory.
• Tools: function calling for multi-step agent tasks, plus ready-made Ollama launchers for Claude Code, OpenCode, Hermes Agent and OpenClaw.
• Benchmarks (NVIDIA’s claims): the #1 open model in its class on PinchBench and #2 open model on BrowseComp.

Bolt vs. the other Nemotron 3.5 models

Model	Size	Built for	Availability
Nemotron 3.5 Super VL	120B parameters	Always-on enterprise agents on servers and DGX systems	Out now
Nemotron 3.5 Bolt 74B-A9B	74B total, 9B active, about 24GB	Local agents on RTX PCs, RTX Spark laptops and DGX Spark	Out now
Nemotron 3.5 Bolt E4B	Fits in 4GB	Fast voice and conversational agents, Jetson and RTX	Due October 23, 2026

Is Nemotron 3.5 Bolt free?

Yes. The weights are free to download from Hugging Face under the OpenMDW-1.1 license, and the model card says Bolt is ready for commercial or non-commercial use. There’s no subscription and no per-token fee when you run it on your own machine. Your real costs are the hardware and electricity.

NVIDIA’s announcement lists Bolt on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM. It doesn’t list a hosted NVIDIA API for Bolt, so plan to run it yourself.

Nemotron 3.5 Bolt Pricing (2026)

Way to get it	Price	What you download
Hugging Face (NVIDIA)	Free	24GB packed 2-bit model, plus a BF16 version for fine-tuning
Ollama	Free	nemotron-3.5-bolt, about 25GB
Unsloth GGUF	Free	IQ2_XXS quant, about 22GB
Your hardware	Varies	An NVIDIA GPU or system with at least 24GB of memory

Nemotron 3.5 Bolt system requirements

• NVIDIA hardware: NVIDIA says Bolt is built for RTX Spark laptops and also runs on GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations.
• Memory: the packed model is about 24GB. Unsloth recommends budgeting around 32GB of RAM or unified memory for short-context use, and 48 to 64GB for more headroom. Long contexts, images and video need extra.
• Mac: Ollama says Apple Silicon and Metal are not currently supported. Unsloth’s desktop app runs on macOS, and a community MLX conversion exists, but neither is an official NVIDIA path.
• Servers: the model card lists vLLM on NVIDIA Ampere (A100), Hopper (H100 80GB) and Blackwell GPUs.

How to run Nemotron 3.5 Bolt

With Ollama (easiest)

Install Ollama, then run ollama run nemotron-3.5-bolt. The Ollama model page lists a 25GB download and 256K context. To use it as the brain of a coding agent, run ollama launch claude --model nemotron-3.5-bolt for Claude Code or ollama launch opencode --model nemotron-3.5-bolt for OpenCode.

With Unsloth or llama.cpp

Unsloth had GGUF files ready on launch day. In Unsloth Desktop you search for Nemotron 3.5 Bolt and download the IQ2_XXS quant, or serve it from the command line with unsloth run --model unsloth/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-GGUF:IQ2_XXS --ctx-size 4096. Start with a 4K context and raise it as your memory allows. Unsloth also runs a small draft model by default for faster output, which needs 1 to 2GB more memory.

Our tip: if Bolt crashes or slows to a crawl, lower the context size before anything else. On a 24GB card, a long context and image inputs are what push it over the edge.

Limitations

• NVIDIA-first: official instructions are for CUDA, and Ollama doesn’t support Apple Silicon yet.
• Tight fit on 24GB: the model alone takes about 24GB, so long chats or images may not fit on a 24GB card.
• 2-bit trade-off: the aggressive quantization is what makes it fit, and NVIDIA positions Bolt as smart enough for most jobs, not as its smartest model.
• Text output only: it can read images and video but doesn’t generate them.
• Self-reported scores: the PinchBench and BrowseComp rankings come from NVIDIA.

Nemotron 3.5 Bolt alternatives

• Other local models: compare sizes and hardware needs in our guide to the best local AI models.
• Reflection Beam: a much larger open-weight coding model. See Reflection Beam and our guide on running Reflection Beam locally to see why it needs far more hardware than Bolt.
• Nemotron 3.5 Bolt E4B: if 24GB is too much, NVIDIA’s 4GB version is due on October 23, 2026.

Verdict

Nemotron 3.5 Bolt is one of the most practical open models for running an AI agent on your own PC: free weights with a commercial-friendly license, a one-line Ollama install and a 24GB footprint. It’s a great fit if you have an NVIDIA card with 24GB or more and care about privacy. Mac users and anyone on a smaller GPU should wait for broader support or the 4GB E4B model.

Nemotron 3.5 Bolt brings a capable, free agent model to a single 24GB NVIDIA GPU, with a license that allows commercial use and day-one support in Ollama and Unsloth.

Try it if you have the hardware; Mac users and smaller GPUs should watch for the 4GB E4B model.