NVIDIA released two new open-weight models on October 15, and they're aimed at quite different machines. Nemotron 3.5 Super VL is the big one, meant for servers. Nemotron 3.5 Bolt is a compressed version of it that NVIDIA built to run agent tasks locally.
Super VL is a mixture-of-experts model with 120B total parameters and 12B active per token. It mixes Mamba-2 layers with Transformer attention, and it adds a vision encoder, so it takes text, images, documents and video (clips up to two minutes) and answers in text. The context window goes up to 1M tokens. NVIDIA published BF16 weights, whose reference setups call for four H100s or two GB200s, plus an NVFP4 version for DGX Spark and data center GPUs. It runs on vLLM and SGLang, and you can switch its reasoning mode on or off in the chat template.
Artificial Analysis tested it before launch and gave it 26 on its Intelligence Index, which by its count makes Super VL the leading US open-weights model right now. That is double the 13 that Nemotron 3 Super scored, and Artificial Analysis says the new model uses about 60% fewer output tokens per task. On a pre-release DeepInfra endpoint serving the NVFP4 weights, it measured about 630 tokens per second with 100K-token prompts.
Bolt is the one more people will be able to try at home. NVIDIA took Super VL and shrank it to 74B total parameters with 9B active, then quantized all of its experts to 2 bits during training, which the model card calls "a native 2-bit quantization to fit in 24GB of memory." NVIDIA says it tuned the compression so the model stays under 24GB at batch size 1 with a 128K context, and the model itself supports up to 256K. It still takes images and video. The card lists vLLM on Linux with deployment settings for A100, H100 and Blackwell GPUs, and the hardware NVIDIA says it tested on includes a GeForce RTX 5090 and DGX Spark.
Artificial Analysis puts Bolt at 21.4, fourth among models under 100B parameters it has tested and first among US models in that group, though it notes Bolt uses about 1.8 times as many output tokens per task as Super VL. In NVIDIA's own model card, Bolt beats Qwen3.6-35B-A3B and Gemma 4 26B-A4B on most of the agent and reasoning benchmarks it lists.
Both models are under OpenMDW-1.1, a short permissive license. You can use, modify and sell products built on them without paying anything. If you redistribute the weights, you have to include the license and keep the original notices. Your rights end if you sue someone claiming the model infringes a patent or copyright, and the license puts no limits on what you do with the outputs.