NVIDIA opens more multimodal AI on its machines
Super VL and Bolt target different hardware budgets, with weights now published.
Weights published; E4B announced for October 23
Facts, analysis and limits
Verified facts
- Super VL: 120 billion parameters, 12 billion active and context up to a million tokens; Mamba-2 hybrid architecture, MoE and attention with visual encoder.
- Bolt 74B-A9B: native 2-bit quantization, an advertised 24 GB footprint and up to 256K context; actual requirements depend on context length and concurrency.
- The model cards date both releases to October 15 and state global distribution under OpenMDW-1.1.
- The Super VL vLLM recipe for DGX Spark serves one concurrent sequence; the model's maximum context window does not guarantee a given throughput on every machine.
- NVIDIA claims up to 5× the throughput of models in the same class; this journal has not reproduced the measurements. Bolt E4B 4 GB has not yet been released in this announcement.
What changes
Deployment control becomes more concrete, but open means neither free to operate nor universally compatible.
Practical implication
Start from your use case and available memory; compare total costs with cloud hosting without treating claimed throughput gains as a guarantee.
Availability
Public weights and model cards; global scope according to NVIDIA, with no Belgium/EEA exclusion mentioned. Size local hosting appropriately; check API providers, costs and regions separately.
Limits and uncertainty
Documentary publication confirmed, without downloading the full weights or benchmarks executed. Do not confuse Bolt's 24 GB with the memory required by Super VL, nor advertise E4B available on 15.