NVIDIA Nemotron 3.5: A 74B AI Agent Model That Runs on One GPU

On October 15, 2026, NVIDIA released Nemotron 3.5 — a family of open AI models built for agents. The standout: Bolt 74B-A9B, a 74-billion-parameter vision-language model compressed into a 24GB footprint, so a single high-end GPU can host a real agentic AI.

The three models

• Super VL 120B-A12B — flagship for always-on enterprise agents. 120B total / 12B active (MoE), 1M-token context, reads text, images, video and PDFs. Available now.
• Bolt 74B-A9B — 74B total / 9B active, native 2-bit quantisation (IQ2), 256K context. Fits in 24GB. Available now.
• Bolt E4B — 4GB featherweight for conversational agents, voice assistants and physical AI. Due October 23, 2026.

What the numbers say

Super VL: NVIDIA claims it matches its largest reasoning model at ~1/5 the parameter count and completes agent tasks up to 5x faster than the next-best open model in its class — its own benchmark, not independently verified. Its model card is more careful: it beats Qwen3.5-122B on most agent tests but trails GLM-5.3-Flash and DeepSeek V4 Flash on coding and browsing.

Bolt 74B on NVIDIA's card, vs Qwen3.6-35B-A3B:

• BrowseComp: 66.75 vs 48.74
• Terminal-Bench 2.1: 50.98 vs 44.38
• GPQA Diamond: 86.43 vs 83.40
• PinchBench: 87.82 vs 88.07 — Bolt loses this one, despite the announcement calling it "#1"

Why local agents matter

An agent that browses, codes and debugs on your own machine keeps your documents off the cloud entirely — for law firms, clinics and compliance-heavy teams, "the data never leaves the device" beats any benchmark point. NVIDIA's test hardware includes a PC with a GeForce RTX 5090 and DGX Spark; RTX Spark laptops went on sale the day after release from $2,599.

License and how to run it

Weights, datasets and training recipes are open under OpenMDW-1.1 (permissive, commercial use allowed, keep licence and notices on redistribution). Bolt runs on the standard local stack — llama.cpp, Ollama, LM Studio, vLLM, Unsloth — with DFlash speculative decoding in the supplied recipes. NVIDIA says it sustains at least 80 tokens/sec.

Caveats: 24GB still means a high-end GPU; "local" isn't "private" without proper access controls; and the small print on the licence's patent/copyright termination clause is worth reading before you bet infrastructure on it.

---

This is a condensed explainer. The full breakdown — the three-model comparison, a 3-step local setup guide, the licence details and the honest caveats — is on NeurAITool: NVIDIA Nemotron 3.5: Run AI Agents on Your Own PC. Canonical: https://www.neuraitool.com/2026/10/nvidia-nemotron-35-open-ai-models-that.html