NVIDIA Releases 120B Nemotron 3.5 Super VL and Bolt Models for Local Agents

NVIDIA has covered two tiers of the agent stack at once: server and desktop. The company released Nemotron 3.5 Super VL, an open multimodal 120B-parameter model for specialized agents that run nonstop, along with a pair of Bolt models for local deployment. Bolt 74B-A9B 24GB targets RTX and DGX Spark, while Bolt E4B 4GB, which NVIDIA pitches for fast conversational agents, fits in 4 GB. Together with Lightning (high throughput, released in August) and Ultra (the hardest reasoning tasks), the Nemotron family now spans the full range, from enterprise servers to a laptop.

Super VL: big-model accuracy with one-fifth the parameters

NVIDIA's main pitch is economics. Super VL, built together with the Nemotron Coalition, matches NVIDIA's own largest reasoning model in accuracy on multimodal agent tasks (instruction following, tool calling, multistep business workflows, image understanding) while having roughly five times fewer parameters. On the Artificial Analysis Intelligence Index leaderboard, the company says the model delivers up to 5x higher throughput than other models in its class. For agents whose tasks run for days or weeks, that translates directly into a lower cost per token and shorter execution time.

Under the hood: about 26T tokens across pretraining and continued pretraining, followed by roughly 3.8M post-training examples and reinforcement learning in dozens of environments. According to the company, NVIDIA Dynamo can serve up to 8x more users at 85% lower latency. The vision side is tuned for routine office work: OCR, tables, charts, and structured data in PDFs and on web pages. A single model instead of a text model paired with a vision model means context is not lost on the way from document to data to decision.

The release is open: weights, datasets, and training recipes. Hugging Face hosts datasets for agentic tool use, finance, instruction following, long context, fairness, and vision-language, and the RL environments are available in NeMo Gym.

Bolt: 74 billion parameters in 24 gigabytes

Bolt 74B-A9B 24GB is built around native 2-bit quantization. The back-of-the-envelope math: 74B weights at 2 bits each come to about 18.5 GB, so the model stays within 24 GB and leaves room for other applications. Judging by the A9B designation, only a fraction of the parameters, about 9B, is active per token, which accounts for the speed. NVIDIA positions the model as "smart enough to handle most tasks and noticeably faster than the smartest models" at the same memory footprint. The claimed rankings among open models are #1 in its class on PinchBench and #2 on BrowseComp.

The intended use is delegating a project to a local agent harness: web research, writing, code debugging, with no data sent to the cloud. Support is announced for llama.cpp, Ollama, LM Studio, Unsloth, Canonical, and vLLM. The model runs on GeForce RTX, DGX Spark, DGX Station, RTX PRO, and RTX Spark laptops.

Bolt E4B 4GB is optimized for intelligence per gigabyte and aimed at physical and conversational agents on GeForce RTX, Jetson, and RTX/DGX Spark: voice assistants, interactive characters, and embedded interfaces.

Who is already fine-tuning Super VL

NVIDIA names dozens of organizations that are fine-tuning the model on their own data, tools, and policies:

• Scale AI: a supplier credit financing agent for Skydance and an M&A due diligence agent for EY.
• AfterQuery: credit agreement analysis for Hebbia.
• AI21 and Alice: cyberattack resilience.
• Mercor: coding agents (planning, writing, debugging), trained on the APEX-SWE benchmark.
• ModelScope: fine-tuning through the ms-swift framework.
• Ramp: intelligence for financial documents.
• Dream: sovereign AI for cybersecurity and vulnerability analysis.
• Lila Sciences: scientific search and discovery.
• R1: medical insurance claims analysis and revenue cycle management.
• TrainLoop: fine-tuning on dermatology data from Nolla, the first company authorized to use AI to draft initial prescriptions.

Among platforms, integrations have been announced by Atlassian (agentic search across Jira and Confluence), CrowdStrike (agentic blue teaming), Dassault Systèmes (engineering design and simulation), Perplexity (Search and Computer), Salesforce (government process automation through Missionforce Operations), and Synopsys (semiconductor design verification working together with the model).

The practical takeaway for teams is this. If data cannot leave the perimeter (insurance, credit scoring, healthcare), Super VL can be deployed on-prem and fine-tuned through NeMo, and NIM provides a ready-made microservice with updates every two weeks. If you need a private agent on a workstation, you can test Bolt 74B-A9B on 24 GB of VRAM in an evening through Ollama or LM Studio. For embedded voice scenarios, E4B on Jetson is the fit. NVIDIA says teams no longer have to choose between capability and deployability, but real task-level numbers are worth measuring on your own data. Super VL is available on Hugging Face, OpenRouter, the ModelScope community, and as NVIDIA NIM.