ZDNET’s key takeaways
- Three new Nvidia open models strive for efficiency and speed.
- Open models are having a renaissance on the heels of the Hugging Face hack.
- US labs are catching up in the open model race.
China-based AI labs have had the upper hand on competitive open-weight AI models over the last few years, but a recent surge in Western-made open models is challenging that.
Also: Open weights vs. closed: An AI civil war’s afoot
Nvidia’s new model, the open-weight Nemotron 3.5 Super VL, aims to be the latest data point in that surge. The 120-billion-parameter model, available today, targets the industry’s increasing focus on agentic work by delivering “the fastest task completion and highest accuracy for specialized, always-on agents,”
“As the fastest open agentic model in its class — 5x faster than the next best — it helps enterprise agents complete high-volume tasks faster across workflows that span days or weeks,” Nvidia said in the release. That’s based on Artificial Analysis’ benchmark testing of the model against its similarly sized competitors; InclusionAI’s Ling 3.0 Flash and DeepSeek’s V4.1 Flash came in second and third, respectively, but with significant distance between them and Super VL.
Nvidia also said Super VL achieves top-tier accuracy for specialized agents with post-training. The company cited how far post-training a base version of the model could take its capabilities — in one case, a 44% gain in medical insurance claims, which put the model far above Claude Opus 5. While that means the base model didn’t top proprietary frontier models on its own, it does give enterprises the chance to tailor the model to their needs and see highly specific gains.
As Nvidia pointed out in the release, the advantage open models have over proprietary models is that businesses can post-train them on their own data and run them on their own infrastructure, offering greater personalization and security.
Super VL is flanked by two new edge models as well: Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB (no, Nvidia didn’t waste any time on naming here). Edge AI refers to models that can operate locally and off-grid. Nvidia said the former Bolt is built for local agentic use, while the latter Bolt is for “fast, conversational agents.”
US open-weight AI takes off
While it’s unclear whether Nvidia’s new model will raise the bar for open-weight or proprietary models in practice, Artificial Analysis’ findings demonstrate efficiency gains for now.
Also: I tested Gemini Spark – another Google AI fail, or its next dark horse?
In the bigger picture, however, the release is part of a broader shift in the AI world. Open models from Chinese companies like DeepSeek and Moonshot.ai have historically embodied the promise (and peril) of open, cheaper, more customizable models compared with closed, expensive ones. But French lab Mistral just launched its Le Chonk model, which hit certain state-of-the-art benchmarks. Reflection AI, which Nvidia has invested in and is in talks to acquire, released its Beam model last week, also to much fanfare.
The geography gap may be closing in the open-weight wars.
The industry is also embracing open-weight models more than it used to, following the OpenAI Hugging Face breach, which dispelled the previous notion that only open models posed security risks. If US-made open models prove themselves against Chinese ones, that could help open models take off in areas where they haven’t yet, given cybersecurity guidance against using foreign tech.
