Nemotron 3.5: NVIDIA shrinks its agentic model until it fits on a 24 GB gaming GPU

The announcement, signed by Kari Briski on NVIDIA's official blog, introduces Nemotron 3.5 Super VL as a multimodal model with 120 billion total parameters, 12 billion of them active per token. According to NVIDIA, it combines reasoning and tool use with an understanding of text, images, documents and video, and it is built for agents that work with charts, web pages and documents. The company calls it "the fastest open agentic model in its class", roughly five times faster than the best competitor, and claims accuracy comparable to its own largest reasoning model with about a fifth of the parameters. This is where it pays to slow down: the blog names neither the competitor nor the internal reference model, and it publishes no scores. These are vendor claims, not comparable measurements. The model is available on Hugging Face, OpenRouter, ModelScope and as a NIM microservice; the stated training covers about 26 trillion tokens across pre-training and continued pre-training, around 3.8 million post-training samples and reinforcement learning across dozens of environments, with open datasets released and RL environments in NeMo Gym.

The interesting part is the second model. Bolt 74B-A9B 24GB has 74 billion total parameters and 9 billion active, and thanks to native 2-bit quantization it fits in 24 GB of memory: it runs on GeForce RTX PCs, RTX Spark laptops, DGX Spark, DGX Station and RTX PRO workstations. The Hugging Face model card fills in the technical details: a hybrid Mamba-2/Transformer architecture called LatentMoE, a C-RADIO v4-H vision encoder, text, image and video input, a 256K-token context window and the OpenMDW-1.1 license. It also explains where the model comes from: it is Super VL 120B-A12B compressed with NVIDIA's Star Elastic method (arXiv 2605.07182), with active experts cut from 22 to 9 and a 120-billion-token post-compression recovery phase.

For Bolt, unlike Super VL, NVIDIA does publish numbers in the model card. The IQ2 quantized version scores 86.43 on GPQA Diamond, 50.98 on Terminal-Bench 2.1 and 66.75 on BrowseComp, set against Qwen3.6-35B-A3B, Gemma4 26B-A4B and Nemotron 3.5 Lightning 30B-A3B. It doesn't win everywhere: on AA-Omniscience accuracy it scores 16.90 against Gemma4's 22.17, and on PinchBench 87.82 against Qwen3.6's 88.07. NVIDIA lists support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM, the tools people actually use to run a model locally. The third model, the 4 GB Bolt E4B, is aimed at conversational agents, voice assistants and interactive characters, including on Jetson, and NVIDIA says it will be available on October 23 — an announced date, not a release that has happened. Among early adopters the company names Scale AI (with versions for Skydance and EY), AfterQuery (for Hebbia), AI21, Alice and Mercor, and it announces integrations with Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys. With NVIDIA Dynamo it claims up to 8 times more users served and 85% lower latency, again without saying compared to what.

What stays with me is the asymmetry between the two announcements inside the same announcement. For Bolt we know the license, architecture, context, benchmarks (reported by NVIDIA) and even where it loses. For Super VL, the bigger model and the one the speed claims rest on, I couldn't open the Hugging Face model card because access was blocked, so I haven't verified its license or context length — an NVIDIA forum mentions one million tokens, and an Artificial Analysis index score of 26.5 circulates on secondary sites, but I couldn't find it at the source. The 2-bit quantization that squeezes a 74-billion-parameter model into a 24 GB consumer GPU is still the technical story: if those 24 GB hold up under real workloads, it changes who can run a multimodal agent at home. But that is for independent measurements to tell, and so far there are none for any of the three models (and Bolt E4B isn't out yet).

— Olya