Nvidia’s new Nemotron 3.5 Super VL is another step forward for US open-weight AI

Nvidia has released a new open-weight model, Nemotron 3.5 Super VL, which the company says is available today. The model has 120 billion parameters and is aimed squarely at agentic workloads, the class of applications where a model runs tasks over extended periods rather than answering a single question. Nvidia describes it as delivering the fastest task completion and highest accuracy for specialized, always-on agents.
Two smaller models ship alongside it, both aimed at edge deployment, meaning they can run locally without a network connection. The first is Nemotron 3.5 Bolt 74B-A9B 24GB, which Nvidia positions for local agentic use. The second is Nemotron 3.5 Bolt E4B 4GB, intended for fast, conversational agents. The naming encodes the memory footprint each model is designed around, so a 24GB model fits comfortably on a higher-end workstation GPU, while a 4GB model can run on much more modest hardware.
Nvidia claims both Bolt models deliver the highest intelligence per gigabyte of memory relative to its other models. The company says the 24GB version achieves the best split of capability and speed per gigabyte, while the 4GB version delivers the best intelligence per gigabyte for its intended use cases, which Nvidia lists as voice and device agents and interactive characters, among others.
The headline performance claim is a speed comparison. Nvidia says Super VL is the fastest open agentic model in its class and five times faster than the next best, citing benchmark testing by Artificial Analysis against similarly sized competitors. Artificial Analysis is an independent outfit that runs standardized evaluations across models, so its numbers are at least comparable across vendors in a way that vendor-reported benchmarks often are not. According to that testing, InclusionAI's Ling 3.0 Flash placed second and DeepSeek's V4.1 Flash third, with Nvidia describing a significant gap between those two and Super VL.
The accuracy story is more nuanced. Nvidia says Super VL reaches top-tier accuracy for specialized agents after post-training, which is the process of continuing to train a base model on domain-specific data after its initial general-purpose training. The company highlights how far post-training a base version of the model can move its capabilities, citing a 44% gain in medical insurance claims that put the tuned model well above Claude Opus 5. Read carefully, that framing concedes something important: the base model did not beat proprietary frontier models on its own. The value proposition is that a business can take the open weights and tune them on its own data to get large, narrow gains in its own domain, rather than relying on a general model's zero-shot abilities.
That is the core argument for open weights in enterprise settings, and Nvidia makes it explicitly in the release. Businesses can post-train open models on their own data and run them on their own infrastructure, which the company frames as offering greater personalization and security than closed alternatives. In practice, a team that already has labeled domain data, such as claims records or support transcripts, can capture the gains Nvidia describes without sending that data to a third-party API. Teams without such data, or without the compute to fine-tune a 120-billion-parameter model, will see less of the benefit.
The release lands amid a broader shift. Chinese labs including DeepSeek and Moonshot.ai have historically dominated competitive open-weight releases, offering cheaper and more customizable alternatives to closed, expensive models. Recent Western entries are challenging that: French lab Mistral launched a model called Le Chonk that hit certain state-of-the-art benchmarks, and Reflection AI released a model called Beam last week, drawing considerable attention. Nvidia has invested in Reflection AI and is reported to be in talks to acquire it. The effect, read broadly, is that the geographic gap in open-weight AI may be narrowing.
There is also a security dimension worth noting. The OpenAI Hugging Face breach, as it is being described, undermined the earlier assumption that only open models carried security risk from public weights. Industry posture toward open-weight models has loosened as a result. For US enterprises, a domestically produced open model may also clear procurement hurdles where cybersecurity guidance discourages reliance on foreign technology, which could matter in regulated sectors.
What remains genuinely uncertain is whether Super VL shifts the practical bar for open or proprietary models. The Artificial Analysis results show efficiency gains at the time of release, but benchmark leadership in this space has historically been short-lived, and the gap between a strong base model and a post-trained specialized one is exactly where most enterprise value is realized or lost. A likely trade-off for adopters is that the specialized accuracy Nvidia highlights depends on doing that post-training work, which is a real engineering investment rather than an out-of-the-box capability.
Why it matters: teams building always-on agents now have another US-produced open-weight option, including a 24GB edge model that can run on local hardware, which matters for latency, offline operation and data residency. The 4GB model lowers the floor for voice and device agents embedded in products. Developers evaluating this against DeepSeek or other open alternatives should test on their own workloads, since the strongest numbers cited are post-training results rather than base-model performance, and those depend on data and compute that not every team has.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Nvidia’s new models join a wave of open-weight releases from US labs on the heels of the Hugging Face hack.