Nvidia squeezed a 74B agent model into 24GB so it can run on your laptop

Nvidia rolled out three new Nemotron models on Thursday, and the clearest headline is the one meant for laptops: Nemotron 3.5 Bolt 74B-A9B 24GB. The company says it can run coding, research, and conversational workloads locally, with enough room left for the kind of multistep reasoning, tool use, and web browsing Nvidia wants to bundle into “everyday agentic tasks.”

The bigger sibling is Nemotron 3.5 Super VL, a 120-billion-parameter vision-language model aimed at enterprise agents. Nvidia says it can handle image and document understanding alongside reasoning and tool use, and that it is built to deliver strong multimodal performance at about one-fifth the parameter count of Nvidia’s largest reasoning model. Unlike Nemotron 3 Super, which launched in March and only took text, Super VL also accepts images and video input.

The benchmark story is more mixed than the marketing. Nvidia says Super VL beats Alibaba’s Qwen3.5-122B-A10B on scientific reasoning and terminal tasks, but falls behind on SWE-bench Verified and document understanding. On Terminal-Bench 2.1, for example, Nvidia reports 66.7 for Super VL versus 48.7 for Qwen3.5, while SWE-bench Verified comes in at 71.1 versus 71.9. Nvidia also says Super VL does better than Ultra on Hermes Bench and LangChain behavioral tests, but its own charts still show it trailing Qwen3.8-Flash-Next and DeepSeek V4 Flash in some speed-and-accuracy tradeoffs.

The engineering is where Nvidia is clearly trying to make the pitch land. Super VL keeps the Nemotron 3.5 family’s hybrid setup, mixing Mamba-2, attention, and mixture-of-experts layers, then adds a vision encoder. It also uses speculative decoding through LiLiCorr, where a smaller model drafts tokens ahead and the main model checks them in batches. Nvidia says that can make token generation up to 2.3 times faster on DGX Spark compared with decoding without speculation.

Bolt is the more practical trick. Nvidia says it uses two-bit quantization to squeeze the 74B model down to 24GB of memory, with 9 billion active parameters out of 74 billion. In Nvidia’s own Terminal-Bench 2.1 tests, it scores 51.0, ahead of Qwen3.6-35B-A3B and Gemma4 26B-A4B. A smaller Nemotron 3.5 Bolt E4B 4GB is due later this month. All three models are available under the OpenMDW 1.1 license on Hugging Face, including commercial use and modification.