Nemotron 3.5 Bolt

Nemotron 3.5 Bolt is NVIDIA’s compact vision-language model for local coding, reasoning, and tool use. It has 74 billion total parameters, with 9 billion active per token, and combines a Mamba-2/Transformer mixture-of-experts backbone with a C-RADIO vision encoder. NVIDIA model card

NVIDIA developed Bolt from Nemotron 3.5 Super through model compression and further training. Its published GGUF uses quantization-aware distillation for IQ2_XXS expert weights, alongside NVFP4 and Q8_0 for other layers.

• Vision: work with images, screenshots, charts, and documents.
• Tools: call functions and use their results in multi-turn tasks.
• Thinking: enable or disable reasoning before the answer.
• Context: supports up to 256K tokens. Usable context depends on available memory; images, conversation history, tool results, and output share that budget.

Compatibility

NVIDIA provides CUDA deployment instructions. Apple Silicon / Metal is not currently supported. Memory requirements depend on context size, visual inputs, and speculative decoding.

License

Released under the OpenMDW License Agreement, version 1.1. See NVIDIA’s documentation for training details and evaluations.