Nvidia launches new open models to run agents in the enterprise and on PCs
- ID
- 37010
- Status
- summarized
- Published
- 19 Oct 2026, 11:37 PM
- Fetched
- 20 Oct 2026, 12:39 AM
- Provider
- TechCrunch
- Category
- technology
- Original URL
- https://techcrunch.com/2026/10/19/nvidia-launches-new-open-models-to-run-agents-in-the-enterprise-and-on-pcs/
- Source URL
- https://techcrunch.com/feed/
Summary
- Score
- 4.0
- Created
- 20 Oct 2026, 12:39 AM
- Tags
- Audience
- developersai_ml_learnersai_agent_userssaas_founders
What happened
Nvidia released Nemotron 3.5 Super VL, a 120-billion-parameter open-weight model for enterprise agents, claiming similar instruction-following and tool-use accuracy to its larger Nemotron 3 Ultra with fewer parameters. It also launched two lightweight "Bolt" models meant to run locally on AI PCs using the RTX Spark chip and on DGX Spark workstations, shipping now through partners including Microsoft and Dell. Nvidia says Super VL finishes tasks faster than Anthropic and OpenAI models (named as Opus 5, GPT 6.1 Sol, Luna, Astra) when post-trained on agentic work like document intelligence, financial accounting, medical insurance claims and security resilience, and lists Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce as users.
Why it matters
The only decision-relevant fact here is a 120B open-weight model pitched at roughly Ultra-level tool-use accuracy, which is a candidate for self-hosted agent inference instead of per-token API billing. But the post gives no license terms, no benchmark numbers, no memory footprint or throughput for the Bolt models, and no price for RTX Spark or DGX Spark hardware, so you cannot size a deployment or compare cost against your current API spend from this text alone. Treat it as a shortlist item to benchmark on your own tool-use tasks, not a migration trigger.
Discussion angle
Nvidia's speed claim rests on models "post-trained on specific agentic tasks" — so which of your agent workloads is narrow and stable enough to post-train on, and what would you need (eval harness, license text, GPU memory) before swapping out a hosted frontier model?