發生什麼事
This release has three models: Super VL plus two edge models, Bolt 74B-A9B 24GB and Bolt E4B 4GB (NVIDIA blog). Super VL activates 12 billion parameters at a time. Give it text, images, documents or video and it can make sense of them, and it can also reason and call tools (@NVIDIAAI). According to NVIDIA, Super VL is only about a fifth the size of the company's biggest reasoning model, yet the two are on par for accuracy in multimodal agent tasks; on throughput it leads its class, by as much as five times. Training recipes, datasets and weights are all public, and the model is on Hugging Face and other platforms. Native 2-bit quantization is what lets Bolt 74B-A9B get by on 24GB of memory; it can run agent work locally on RTX PCs and DGX Spark and supports Ollama and LM Studio. Bolt E4B, arriving October 23, targets conversational agents such as voice assistants (NVIDIA blog).
01Sources
- Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models: official NVIDIA blog; model specs, NVIDIA's own performance statements, availability on platforms
- @NVIDIAAI: NVIDIA AI's official account introducing Super VL, parameter counts
- Hugging Face model page: Super VL open weights