Nvidia launches three new HuggingFace models, including Nemotron‑Nano‑9B‑v2 and

Nvidia’s latest push into open‑source AI models arrives with three fresh uploads to Hugging Face, the first of which—Nemotron‑Nano‑9B‑v2—has already logged more than 200 000 downloads in its debut week, according to the model’s Hugging Face page. The 9‑billion‑parameter transformer is positioned as a multilingual text‑generation engine, supporting English, Spanish, French, German, Italian, Japanese and two additional languages, and is built on Nvidia’s Nemotron‑Nano architecture. The repository lists a suite of training datasets, including the Nemotron‑Post‑Training‑Dataset‑v1 and v2, the Nemotron‑Pretraining‑Dataset‑sample, and the Nemotron‑CC‑v2 and CC‑Math‑v1 corpora, indicating a blend of general‑purpose and domain‑specific material (source: Nvidia on Hugging Face). With 479 “likes” and a rapid download rate, the model’s early traction suggests developers are eager for a compact, GPU‑optimized alternative to larger, cloud‑only offerings.

The second model, fourcastnet3, is a weather‑forecasting transformer that builds on a series of recent arXiv pre‑prints (2507.12144, 2402.16845, 2408.03100, 2408.01581, 2306.03838) cited in its metadata. Although its download count is modest—280 in the first week—the model’s inclusion of an Apache‑2.0 license and a “region:us” tag signals Nvidia’s intent to make the tool readily reusable for U.S. research institutions and commercial partners. The modest uptake reflects the niche nature of high‑resolution climate modeling, but the open‑source release aligns with Nvidia’s broader strategy of seeding the ecosystem with domain‑specific AI assets that can be accelerated on its GPUs.

The third addition, corrdiff‑cmip6‑era5, targets climate‑model intercomparison by learning the statistical differences between CMIP‑6 outputs and ERA5 reanalysis data. Its metadata references a single arXiv paper (2309.15214) and, like fourcastnet3, is tagged “region:us.” With only 75 downloads so far, the model is clearly in an early adoption phase, but its presence on Hugging Face underscores Nvidia’s commitment to providing research‑grade tools for the Earth‑system science community, a sector that has increasingly turned to AI for bias correction and downscaling.

Beyond the three models, Nvidia has also released NV‑Generate‑CT, a latent‑diffusion model for medical‑imaging CT synthesis, and a handful of other specialized assets. NV‑Generate‑CT has attracted 14 downloads and 12 likes, reflecting the high barrier to entry in regulated medical AI. While the download figures are low, the model’s inclusion of an “other” license and a reference to arXiv:2508.05772 suggests Nvidia is testing the waters for open‑source diffusion in clinical contexts, where data privacy and validation remain paramount.

The flurry of releases comes as Nvidia’s market valuation recently breached the $4 trillion threshold, a milestone highlighted by Forbes, which warned that the “real AI boom hasn’t started yet” despite the company’s soaring stock price. Analysts at CES 2026 noted that AI announcements, including Nvidia’s model rollouts, are reshaping the hardware‑software stack, moving the industry away from pure consumer GPU hype toward enterprise‑grade, task‑specific models. By publishing ready‑to‑run transformers on a public hub, Nvidia is effectively lowering the friction for developers to experiment with its hardware, potentially accelerating the adoption curve for both generative text and scientific‑AI workloads.

In practice, the Nemotron‑Nano‑9B‑v2’s multilingual capability could give startups a lightweight alternative to larger models like GPT‑4, especially when deployed on Nvidia’s own H100 or A100 GPUs. The model’s training on both general‑purpose and math‑focused corpora (Nemotron‑CC‑Math‑v1) hints at a design that balances fluency with numerical reasoning, a niche that many enterprise applications—such as automated report generation and multilingual customer support—require. With the open‑source community already showing strong interest, Nvidia’s strategy appears to be less about selling a single monolithic AI platform and more about seeding a diverse ecosystem of specialized models that can be fine‑tuned and scaled on its hardware, a move that could cement its role as the de‑facto infrastructure provider for the next wave of AI innovation.