Nvidia expands Nemotron family with Super VL and two Bolt models
Open weights now cover servers down to 4GB edge devices
The facts
- Nemotron 3.5 Super VL is a 120-billion-parameter open multimodal model, available now on Hugging Face, Open Router, ModelScope and as NVIDIA NIM
- Nvidia says Super VL matches its largest reasoning model's accuracy at roughly one-fifth the parameter count and is 5x faster than the next-best open model in its class on the Artificial Analysis Intelligence Index
- Bolt 74B-A9B 24GB uses native 2-bit quantization to run locally on RTX/DGX Spark within 24GB of memory, and ranks #1 open model in its class on PinchBench and #2 on BrowseComp, per Nvidia
- Bolt E4B 4GB, for conversational and physical agents on GeForce RTX, Jetson and RTX/DGX Spark, arrives Oct. 23; Bolt 74B-A9B is available now, with Nvidia saying it runs 8x more users at 85% lower latency on its Dynamo serving stack
- Super VL was trained on about 26 trillion tokens plus roughly 3.8 million post-training samples and reinforcement learning across dozens of environments, and is being post-trained by Scale AI, AI21, Mercor, Atlassian, CrowdStrike, Perplexity, Salesforce and others
Nvidia has released Nemotron 3.5 Super VL, an open 120-billion-parameter multimodal model. Alongside it come two smaller edge models, Bolt 74B-A9B 24GB and Bolt E4B 4GB.
Super VL and Bolt 74B-A9B are available now on Hugging Face, Open Router, ModelScope and as an NVIDIA NIM microservice. Bolt E4B follows on Oct. 23. Together with the earlier Lightning and Ultra models, the four-model Nemotron line now spans data-centre servers down to 4GB edge devices. That gives organisations an open option at every deployment tier, without trading capability for size.
Nvidia says Super VL matches the accuracy of its largest reasoning model on multimodal agentic tasks, at roughly one-fifth the parameter count. Those tasks include instruction following, tool use, multistep workflows and visual document understanding. The company also claims it is 5x faster than the next-best open model in its class on the Artificial Analysis Intelligence Index leaderboard. The model was trained on about 26 trillion tokens across pretraining and continued pretraining. It was then post-trained on roughly 3.8 million samples, with reinforcement learning across dozens of environments. Nvidia is also releasing open datasets covering agentic tool use, finance, instruction following, long context, fairness and vision-language understanding, plus the RL environments in NeMo Gym.
Bolt 74B-A9B 24GB runs on NVIDIA RTX/DGX Spark, using native 2-bit quantisation to fit within 24GB of memory. Nvidia reports it ranks #1 among open models in its class on PinchBench, and #2 on BrowseComp. Bolt E4B 4GB targets conversational and physical agents on GeForce RTX, Jetson and RTX/DGX Spark hardware. Nvidia says running Super VL on its Dynamo serving stack lets organisations serve up to 8x more users, with 85% lower latency.
A long list of companies is already post-training or integrating Super VL, according to Nvidia. They include Scale AI for Skydance and EY agents, AfterQuery for Hebbia, AI21 and Alice for cybersecurity, and Mercor for coding agents. Platform integrations come from Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys. Nvidia frames the open weights, open datasets and NeMo tooling as giving organisations control over where the model runs and what data it touches. It points to on-premises use cases in insurance fraud screening and credit-score review.
Why it matters
Shipping open weights at three sizes, with matching datasets and RL environments, is Nvidia's bid to make its own hardware the default home for agentic deployment. That goes beyond just selling GPUs. Anyone evaluating closed frontier APIs for document-heavy agent work now has an open, post-trainable alternative. It already has named enterprise adopters in production.
Primary source
NVIDIA — blogs.nvidia.comWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.