NVIDIA Nemotron 3.5 Super VL
NVIDIA’s open-weight vision-language model supports coding, visual reasoning, document understanding and agents that use tools.
NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB
research.nvidia.com
A downloadable vision-language model for local agents, combining quantized weights, configurable reasoning, and tool calling.
Visit website
Pricing
Free
Available on
Linux, Windows
Category
Llm Models
What is NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB?
NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB is a downloadable model designed for local agent applications on compatible NVIDIA systems. It uses a hybrid Mamba-Transformer mixture-of-experts architecture with 74 billion total parameters and nine billion active parameters.
Bolt accepts text, images, and video and returns text, including JSON and tool calls. It targets developers integrating a model into coding, research, and other tool-driven applications, with deployment settings selected for their hardware.
Key features
Compressed agent model
Bolt is compressed from Nemotron 3.5 Super VL, with expert weights quantized to approximately two bits to reduce its memory footprint. Other components use different precisions.
Visual and text inputs
The model processes text, JPEG or PNG images, and supported MP4 videos. Visual inference requires the corresponding components and runtime configuration.
Reasoning controls
Chat-template settings enable or disable reasoning, allowing applications to select the mode for each interaction.
Long-context interaction
Up to 256K context tokens accommodate prompts, visual inputs, conversation history, tool results, and generated responses within one shared window.
Customization and decoding options
NVIDIA releases a BF16 customization checkpoint alongside quantized and GGUF weights. Companion LiLiCorr drafters support speculative decoding in the documented deployments.
Who is it for?
Developer building a local coding agent
Evaluate Bolt as the model behind an agent harness for code changes and debugging.
Research team prototyping tool workflows
Connect selected search and retrieval tools and assess multi-step research tasks.
Application developer handling visual questions
Prototype image-based questions alongside text conversations in a local application.
How to use NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB
- Choose a documented llama.cpp hardware recipe and check its memory and context requirements.
- Build llama.cpp with CUDA support for the selected system.
- Download the main GGUF and vision projector; add the companion drafter when using speculative decoding.
- Start the API server with the recipe's cache and concurrency settings, then connect your application.
NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB pricing
Free
The downloadable model permits commercial and noncommercial use under OpenMDW 1.1 without a license charge. Hardware and hosting are separate deployment costs.
Check current pricing ↗What to know before you use it
- The 24GB designation describes the packed model. The GGUF guide's 28GB recipe targets RTX Spark 64GB or discrete GPUs with at least 32GB; its 32GB recipe targets RTX Spark 128GB or discrete GPUs with at least 48GB.
- The vLLM quick start requires a specified nightly build or newer, b12x kernels, and a reasoning-parser plugin. Its software-integration section specifies Linux and Blackwell SM120 or SM121.
- Documented MP4 input is limited to two minutes, with resolution-dependent frame sampling. Tool execution must be implemented by the surrounding application.
Frequently asked questions
Which operating systems does the GGUF deployment support?
The model-specific llama.cpp integration lists Windows and Linux with NVIDIA Ada Lovelace or Blackwell hardware.
Does every deployment provide the full context window?
No. The GGUF guide pairs 65,536 tokens with its 28GB recipe and 262,144 tokens with its 32GB recipe. Both serve one concurrent sequence.
Can the model be customized?
NVIDIA provides the unquantized BF16 checkpoint for further customization, separately from the main quantized inference checkpoint.