In brief
- What changed: NVIDIA released Nemotron 3.5 Super VL and Bolt 74B, open models for AI agents on servers and local PCs.
- Available now? Super VL and Bolt 74B are downloadable now. Bolt E4B arrives October 23. No regional limit stated.
- Price: Open weights; no license fee in the announcement. You need matching NVIDIA hardware for local runs.
- Who cares: Developers and teams building agents on NVIDIA gear, including RTX and DGX Spark.
NVIDIA launched Nemotron 3.5 Super VL, an open ~120-billion-parameter model for agents, plus Nemotron 3.5 Bolt for local runs.
AI Desk · Published October 15, 2026, 8:10 PM IDT
First look: this story is based on NVIDIA’s official announcement and has not been hands-on tested by Maitz.
The post went live on October 15, 2026 on the NVIDIA blog. According to NVIDIA, Super VL and Bolt 74B are available now, and Bolt E4B ships on October 23.
What is new in Nemotron 3.5?
Super VL targets always-on agents: instruction following, tool use, multistep workflows, and visual understanding. NVIDIA says it reaches up to 5x higher throughput than other models in its class on the Artificial Analysis Intelligence Index leaderboard.
| Model | Size / memory | Target | Availability |
|---|---|---|---|
| Nemotron 3.5 Super VL | ~120B parameters | Enterprise server agents | Now |
| Bolt 74B-A9B | 24GB (2-bit quantization) | Local agents on RTX/DGX Spark | Now |
| Bolt E4B | 4GB | Edge voice and physical agents | October 23, 2026 |
Bolt 74B is meant for everyday local work: web research, writing, and debugging without sending jobs to the cloud. NVIDIA says it ranks #1 open in its class on PinchBench.
Open weights are available to download, including on Hugging Face. NVIDIA also released training datasets and NIM as a production microservice.
Why it matters
An open model lets a team post-train on its own data and run on its own stack. That is different from a closed model that only ships through an API.
What we don’t know yet
- Language quality. The announcement does not detail supported languages.
- Exact hardware needs outside the 24GB and 4GB footprints named in the post.
- Cloud NIM pricing versus self-hosted download cost.
What it means for you
- Developers with NVIDIA GPUs: Download Super VL or Bolt and try a local agent.
- Privacy-minded teams: Bolt is built to keep everyday agent work on the machine. Check that you have enough VRAM.
- Without matching hardware: Follow the weights on Hugging Face, or use ready tools from the AI tools directory.
- Local AI PCs: Our RTX Spark story covers Windows PCs built for large local models.
More on Maitz
Sources
- NVIDIA Blog: Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models, October 15, 2026
- Verified: October 15, 2026
