For Developers/Models/Nemotron 3.5 Bolt 74B-A9B
Open SourceNVIDIAReleased 2026-10

Nemotron 3.5 Bolt 74B-A9B

NVIDIA's open 74B multimodal agent model, with a 2-bit build that fits in about 24GB

Context window

256K

Input / 1M tokens

Free

Output / 1M tokens

Free

Provider

NVIDIA

Open weights under the OpenMDW-1.1 license, which allows commercial and non-commercial use; free to download and run. NVIDIA publishes a 2-bit 24GB build for local inference and full BF16 weights for customization. NVIDIA lists no hosted per-token price. · Data verified 2026-11-08

Nemotron 3.5 Bolt 74B-A9B (October 15, 2026) is NVIDIA's open-weight vision-language model for running everyday agentic tasks locally. It is a hybrid Mamba-2 and Transformer Mixture-of-Experts with 74B total and 9B active parameters, built by compressing Nemotron 3.5 Super VL, and it takes text, images and videos up to two minutes long with up to 256K tokens of context. NVIDIA ships a 2-bit (IQ2) build trained with quantization awareness that fits in about 24GB of memory for RTX GPUs and RTX Spark, alongside full BF16 weights meant for fine-tuning.

How to access it

Download from Hugging Face. The 24GB GGUF build (nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF) runs in llama.cpp, Ollama, LM Studio and Docker Model Runner on NVIDIA Ada Lovelace or Blackwell GPUs. The BF16 weights (nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16) serve with vLLM or SGLang on A100, H100 or Blackwell data-center GPUs.

Strengths

  • ✓Runs locally: the 24GB build targets NVIDIA RTX GPUs and RTX Spark
  • ✓Text, image and video input with up to 256K tokens of context
  • ✓Strong agent scores for its size: 70.0% on BrowseComp and 53.9 on Terminal-Bench 2.1 (NVIDIA-reported, BF16)
  • ✓Reasoning can be switched on, off or to medium effort
  • ✓Commercial use allowed under OpenMDW-1.1

Best for developers who...

Local agents on a workstation GPUPrivate document, screenshot and video understandingFine-tuning a compact multimodal agent model

When to choose it (and when not to)

Reach for Nemotron 3.5 Bolt 74B-A9B when...

  • →You want a capable agent model that runs on one workstation GPU
  • →Your data must stay on your own hardware
  • →You need image and video understanding locally
  • →You want a permissive license for commercial products

Look elsewhere if...

  • ✕You need frontier-level coding or reasoning (large hosted models score far higher)
  • ✕Your GPU has less than 32GB of memory (NVIDIA's local recipes start at 32GB for a 65K context)
  • ✕You need audio input
  • ✕You want a pay-per-token API from NVIDIA

How to use it

  • ›Toggle reasoning with enable_thinking in the chat template; add medium_effort for a lighter thinking budget
  • ›NVIDIA recommends temperature 1.0 and top_p 0.95 for text
  • ›Keep video inputs to mp4 clips of two minutes or less
  • ›Use the 24GB build for local inference and the BF16 weights for fine-tuning

Quickstart

Bash
# Run the 24GB build locally with Ollama
ollama run hf.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF

Command from NVIDIA's Hugging Face model card. llama.cpp, LM Studio and Docker Model Runner are also supported.

API model id: nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF

Benchmarks

BenchmarkScoreNotes
GPQA Diamond88.83NVIDIA-reported, BF16; 86.43 for the 24GB build
Terminal-Bench 2.153.93NVIDIA-reported, BF16; 50.98 for the 24GB build
BrowseComp70.00NVIDIA-reported, BF16; 66.75 for the 24GB build
MCP-Atlas55.60NVIDIA-reported, BF16
HLE (no tools)28.73NVIDIA-reported, BF16
GDPval-AA v231.13NVIDIA-reported, BF16

Source: Hugging Face model card (NVIDIA)

Compare Nemotron 3.5 Bolt 74B-A9B

Compare Nemotron 3.5 Bolt 74B-A9B with any other model

Build a comparison →
All model comparisons →

Learn the concepts