Nemotron 3.5 Bolt: Specs, License & How to Run

Nemotron 3.5 Bolt is an open-weight vision-language model from NVIDIA designed for local agentic tasks. Official hosts we opened on 2026-10-17: the ungated BF16 card nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16, the 24GB IQ2 variant nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB, the GGUF sibling nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF, the Ollama model, the technical report PDF, and developer.nvidia.com Nemotron.

This identification page hangs off ChatGPT alternatives and closed LLM APIs. Those mappings are editorial. We have no sourced quality comparison against ChatGPT.

The NVIDIA blog (15 October 2026) calls it a compact vision-language model with 74B total / 9B active parameters and a 256K-token context, derived from Nemotron 3.5 Super through model compression. The 24GB IQ2 variant uses quantization-aware training to fit in 24GB VRAM for local deployment on NVIDIA RTX/DGX Spark. Weights ship under the OpenMDW License 1.1 with commercial use permitted. The blog’s efficiency claims are vendor claims, not a forkalt ranking.

What we verified (2026-10-17)

Field	Value	Source
BF16 HF likes	50	HF API nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16
Card created / last modified	2026-10-15T09:48:07Z / 2026-10-15T10:33:07Z	same
Gated	false	HF API
License (card)	openmdw-1.1	HF cardData.license
LICENSE file	OpenMDW License Agreement, Version 1.1	LICENSE.txt
Pipeline tag	image-text-to-text	HF API
Safetensors params (BF16)	75,962,994,688 BF16 (~76B)	HF safetensors.parameters.BF16
24GB variant likes	46	HF API …-24GB
24GB created	2026-10-15T09:48:49Z	same
Context	262,144 tokens (256K)	NVIDIA docs + HF card
Modalities	text + image + video in → text out	NVIDIA docs + HF card
Architecture (card)	Mamba-2–Transformer hybrid LatentMoE; C-RADIO v4-H	HF card “Model Architecture”
Ollama availability	nemotron-3.5-bolt:74b-a9b, :latest, :74b-a9b-iq2	Ollama library

GGUF and LiLiCorr drafter siblings are also available on the same Hugging Face collection.

On-device claims (vendor / self-reported)

The NVIDIA technical report says the 24GB IQ2 variant is “designed to run everyday agentic tasks locally” and quotes 84 tokens/s on RTX (vendor figure). It uses native 2-bit quantization with quantization-aware training to fit in 24GB of memory and run alongside other applications. The Ollama library shows a 25GB download size for the :74b-a9b tag with 256K context. Those sentences are vendor/self-reported. forkalt has not re-measured speed or memory on a named machine.

How to run (from the official card, not yet tested)

Ollama (simplest):

ollama pull nemotron-3.5-bolt:74b-a9b
ollama run nemotron-3.5-bolt:74b-a9b

vLLM (24GB variant):

vllm serve nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB \
  --served-model-name nemotron-3.5-bolt \
  --trust-remote-code \
  --gpu-memory-utilization=0.85 \
  --allowed-local-media-path=/ \
  --max-num-seqs=16 \
  --limit-mm-per-prompt '{"video": 1, "image": 1}' \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser nemotron_v3 \
  --enable-lora \
  --lora-modules drafter=nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr \
  --speculative-model dflash \
  --speculative-num-draft-tokens 7

The model has configurable reasoning mode controlled by the chat template with enable_thinking=True/False. Recommended sampling copied from the card: temperature 1.0, top_p 0.95 for text. Also available through llama.cpp, LM Studio, and Unsloth. forkalt has not run these commands.

License

LICENSE.txt is titled OpenMDW License Agreement, Version 1.1, Last Updated: October 2026, copyright NVIDIA Corporation. Hugging Face cardData.license is openmdw-1.1.

Paraphrase only, not legal advice:

• Include the copyright notice and this license with copies or substantial portions.
• Prominently display “Built with NVIDIA Nemotron” attribution on products or services that use the model or derivatives.
• Commercial use including Model-as-a-Service is permitted.
• Create derivative works and distribute them under the same license.
• NVIDIA claims no ownership of outputs.

What we did not verify

Vendor benchmark tables on the card and tech report versus other models stay self-reported. The RTX 84 tok/s figure and 24GB memory footprint are upstream. We have no sourced bake-off against ChatGPT or against other local-agent models. We have not loaded the 24GB variant on a named RTX or DGX Spark machine.

Related chat weights on this site: the larger sibling Nemotron 3.5 Super VL, DeepSeek-V4.1-Flash, MiniCPM-V 4.7, and Muse Spark Open on ChatGPT alternatives; LFM3-1.7B (Liquid text 1.73B; not this 74B local-agent VLM — LFM3-1.7B vs LFM3-Omni-2B), Mage-VL-Cowork (Microsoft ~4.7B CUA VLM), Granite Switch 4.1, LFM3 (family; LFM3 vs LFM2.5-24B-A2B), Kolibri-1 and LFM2.5-24B-A2B on closed LLM APIs. Same Nemotron 3.5 family, different job: Nemotron 3.5 Transcribe is offline speech ASR, not this VLM — Nemotron 3.5 Transcribe vs Granite Speech 5.0. Runtimes: vLLM, DeepInfer vs vLLM. Style rewrite is a different job: humanizer.