The week of October 12 was crowded with open-weight releases: Meta’s Muse Spark Open on the 15th, and the same day NVIDIA’s Nemotron 3.5 Super VL. I used AI Weekly’s roundup, which summarizes VentureBeat and others, plus Interesting Engineering. I haven’t run the model. Performance claims are NVIDIA’s.

The basics

  • 120B total parameters, 12B active (a mixture of experts), with a 1M-token context per the VentureBeat headline.
  • Aimed at agents that read PDFs, charts, tables, images and webpages, with reasoning and tool use for long multistep tasks.
  • Trained on roughly 26 trillion tokens, with about 3.8 million post-training samples and reinforcement learning across about 100 open environments.
  • License: OpenMDW, which per the coverage permits commercial use, fine-tuning and redistribution. Interesting Engineering doesn’t state the license, so check the model card.
  • Where: Hugging Face, OpenRouter, ModelScope and NVIDIA NIM, plus a rate-limited free NVIDIA-hosted API. Production use through NVIDIA AI Enterprise reportedly costs $4,500 per GPU per year.

Smaller siblings

  • Nemotron 3.5 Bolt 74B-A9B uses native 2-bit compression to fit in 24 GB of GPU memory, is on Hugging Face now, and runs through Ollama or LM Studio, according to Interesting Engineering. NVIDIA claims it’s first among open models in its class on PinchBench and second on BrowseComp.
  • Bolt E4B has a roughly 4 GB footprint for edge and Jetson use and is scheduled for October 23.

For most readers the 24 GB variant is the interesting one, since the full model is a data-center model. My Mellum 2.1 post covers another small open option.

The claims, and what’s missing

NVIDIA says:

  • Up to 5x faster than a comparable open model, about 625 output tokens per second on four B200 GPUs. The Artificial Analysis leaderboard hadn’t listed the model at launch.
  • Post-trained versions beat Claude Opus 5 and GPT-6.1 Sol on five domain tasks, with gains of 9 to 44% after post-training. Examples: medical insurance claim accuracy from about 28% to 72% (versus about 38% for Opus 5), and security attack resilience from about 35% to 76% (versus about 66% for Sol).
  • It completes 4x more enterprise jobs than its predecessor and uses 123x fewer tokens per correct task.

VentureBeat notes NVIDIA didn’t disclose task definitions, sample sizes or testing protocols. The comparisons also involve models NVIDIA’s partners post-trained for those specific tasks, so “beats Opus 5” means “a specialized version beats it on this task,” not a general ranking.

Inconsistencies

  • OrcaRouter reports a Terminal-Bench 2.1 discrepancy between the model card (66.70) and the technical report (60.67).
  • Tech Insider reportedly says public leaderboards still rank Claude Opus 5.5 and Gemini 4 Argon ahead. I didn’t verify it, but it matches the pattern in my model comparison post: vendor tables lead, independent tables are more modest.

How I’d approach it

  • If you need a vision-and-documents agent you can self-host, it’s worth a test, especially the 24 GB Bolt model.
  • If you want a coding agent model, the evidence here is thin. Run it on your own repo before believing the headline.
  • Read the OpenMDW license text. “Permits commercial use” in a news summary is not the same as reading the terms.

Sources: AI Weekly, Interesting Engineering.