Nemotron 3.5 Bolt
An open-source large language model from NVIDIA. It uses a MoE architecture with quantization-aware training to achieve 2-bit weights, and is designed for local execution on consumer hardware with 24 GB of VRAM, with support for multimodal visual input.
Timeline (last 90 days)
Among these four models, only Nvidia's Nemotron 3.5 Bolt can run within 24GB of VRAM.
Unsloth released the GGUF format on the day Nemotron 3.5 Bolt launched, supporting llama.cpp, Ollama and LM Studio.
Nemotron 3.5 Bolt supports up to 256K tokens of context, accepts text, images, and video as input, and outputs text.
Nemotron 3.5 Bolt deployment goes through vLLM and supports NVIDIA Ampere (A100), Hopper (H100-80GB), and Blackwell platforms.
Nemotron 3.5 Bolt comes with a LiLiCorr drafter for speculative decoding, implemented via DFlash, with 7 speculative tokens by default in the vLLM configuration.
Nemotron 3.5 Bolt uses native 2-bit quantization (IQ2, with quantization-aware training QAT) to compress the model into 24GB of VRAM, allowing it to run on a single consumer-grade or single-card workstation.
On Hugging Face, in addition to the main repository, GGUF versions of Nemotron 3.5 Bolt and a LiLiCorr-GGUF version are also provided.
Nemotron 3.5 Bolt's license agreement is OpenMDW-1.1, allowing commercial and non-commercial use.
The developer of Nemotron 3.5 Bolt is NVIDIA Corporation, the model date range is listed as December 2025 to May 2026, and the post-training data cutoff is May 2026.
Nemotron 3.5 Bolt controls reasoning modes via chat-template, using enable_thinking=True/False to toggle thinking mode, and the recommended sampling parameters are temperature 1.0 and top_p 0.95, both inherited from Super 3.5 VL.
1 more timeline events
All Facts (11)
Among these four models, only Nvidia's Nemotron 3.5 Bolt can run within 24GB of VRAM.
50%UnverifiedNemotron 3.5 Bolt's architecture adopts LatentMoE mixing Mamba-2 and Transformer, and the vision side uses the C-RADIO v4-H encoder.
50%UnverifiedNemotron 3.5 Bolt supports up to 256K tokens of context, accepts text, images, and video as input, and outputs text.
50%UnverifiedNemotron 3.5 Bolt deployment goes through vLLM and supports NVIDIA Ampere (A100), Hopper (H100-80GB), and Blackwell platforms.
50%UnverifiedNemotron 3.5 Bolt comes with a LiLiCorr drafter for speculative decoding, implemented via DFlash, with 7 speculative tokens by default in the vLLM configuration.
50%UnverifiedNemotron 3.5 Bolt controls reasoning modes via chat-template, using enable_thinking=True/False to toggle thinking mode, and the recommended sampling parameters are temperature 1.0 and top_p 0.95, both inherited from Super 3.5 VL.
50%UnverifiedOn Hugging Face, in addition to the main repository, GGUF versions of Nemotron 3.5 Bolt and a LiLiCorr-GGUF version are also provided.
50%UnverifiedNemotron 3.5 Bolt's license agreement is OpenMDW-1.1, allowing commercial and non-commercial use.
50%UnverifiedThe developer of Nemotron 3.5 Bolt is NVIDIA Corporation, the model date range is listed as December 2025 to May 2026, and the post-training data cutoff is May 2026.
50%UnverifiedNemotron 3.5 Bolt uses native 2-bit quantization (IQ2, with quantization-aware training QAT) to compress the model into 24GB of VRAM, allowing it to run on a single consumer-grade or single-card workstation.
50%UnverifiedUnsloth released the GGUF format on the day Nemotron 3.5 Bolt launched, supporting llama.cpp, Ollama and LM Studio.
50%