TL;DR: NVIDIA released two open models on October 15. Nemotron 3.5 Super VL is a 120B-parameter multimodal model (12B active) for business agents. Nemotron 3.5 Bolt 74B-A9B 24GB is a smaller version squeezed to 2-bit so it fits in 24GB of memory and runs agent tasks on your own machine. Bolt works with llama.cpp, Ollama, LM Studio, Unsloth and vLLM. A tiny 4GB model, Bolt E4B, is due October 23.

What happened

NVIDIA announced “Nemotron 3.5 Super VL, an open 120-billion-parameter model” for “specialized always-on agents,” along with two edge models. “Super VL and Bolt 74B-A9B are available now, with Bolt E4B coming Oct. 23.”

According to NVIDIA’s blog, research page and model card:

  • Bolt 74B-A9B 24GB. 74B total parameters with 9B active, compressed from Super VL and quantized to about 2 bits so it fits in 24GB of memory and can “run alongside other applications.”
  • Inputs and context. Text, images and video in, text out, with up to 256K tokens of context.
  • Speed. NVIDIA reports 84 tokens per second on its RTX Spark hardware using a companion speculative-decoding drafter.
  • Where it runs. Hugging Face weights with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM. NVIDIA says it’s built for RTX Spark laptops and also runs on GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations.
  • Use. The model card says it’s “ready for commercial or non-commercial use” under NVIDIA’s OpenMDW-1.1 license.
  • Super VL. Available on Hugging Face, OpenRouter, ModelScope and as NVIDIA NIM.

Why it matters

Running a capable agent model on your own hardware keeps client files and code off third-party servers and removes per-token bills. NVIDIA’s bet, set out on its research page, is that compressing a big model beats training a small one from scratch at the same memory budget, and that 24GB, the memory on many high-end desktop cards, is enough.

The claims are NVIDIA’s own. It says Bolt beats similarly sized models such as Qwen3.6 35B-A3B and Gemma 4 26B-A4B on agent, browsing and terminal benchmarks, and ranks first on PinchBench among models with a similar memory footprint. The 84 tokens-per-second figure is on NVIDIA’s RTX Spark, not on any 24GB card.

What to do

  1. Check your memory. The packed model is 24GB, so it needs a machine with that much GPU or unified memory to spare, plus room for context.
  2. Use the tool you already have. Try it through Ollama or LM Studio if you already run local models; the GGUF checkpoint is for llama.cpp-based tools.
  3. Test it on your real tasks. Run a few of your own agent jobs side by side with your current cloud model before switching.
  4. Need a smaller one? Wait for Bolt E4B (4GB) on October 23 if your hardware can’t spare 24GB.
  5. Running agents in the cloud? Super VL is on OpenRouter, so you can compare it with your current model without new hardware.

Sources & method

Sources

Method. Last checked October 18, 2026. Details and quotes come from NVIDIA’s blog, research page and model card. Benchmark and speed figures are NVIDIA’s own. We haven’t tested these models. No affiliate links on this page. How we work.

Spotted an error? See our corrections policy and tell us.