NVIDIA doubles down: open + proprietary
NVIDIA is pushing both open‑source models (Nemotron family reportedly downloaded 45+ million times) and proprietary solutions — a dual strategy that keeps dev ecosystems broad while monetizing enterprise integrations. The message at GTC: build for both community adoption and paid enterprise support. (blogs.nvidia.com)
NVIDIA named eight inaugural members of its Nemotron Coalition — Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam and Thinking Machines Lab — and said coalition partners will pool data, evaluations and expertise for joint model development. (investor.nvidia.com)) The company said the coalition’s first joint project is a base model co-developed with Mistral AI that will be trained on NVIDIA’s DGX Cloud infrastructure and used as the foundation for the upcoming Nemotron 4 family. (investor.nvidia.com)) NVIDIA’s technical writeup for Nemotron 3 describes a hybrid “Mamba‑Transformer” mixture‑of‑experts architecture designed to push accuracy-to-inference‑throughput for agentic workloads. (arxiv.org)) NVIDIA’s published benchmarks and third‑party coverage state Nemotron 3 Nano delivers roughly 4x higher token throughput than Nemotron 2 Nano, a claim the company ties to Mixture‑of‑Experts and other architectural changes. (edge-ai-vision.com)) All Nemotron training recipes, datasets and evaluation tooling are published in NVIDIA’s NeMo developer hub and GitHub repository, with companion packages (NeMo Gym and NeMo Evaluator) intended for post‑training validation and safety testing. (github.com)) Individual Nemotron assets on Hugging Face show high single‑asset distribution — for example, an FP8 Nano 30B Nemotron build lists over 1.4 million downloads on the platform — while NVIDIA points users to the Hugging Face collection for model access. (huggingface.co)) NVIDIA’s public filings and SEC disclosures indicate multi‑year cloud commitments totaling roughly $26 billion as of the Form 10‑Q covering the quarter ended October 26, 2025, a figure widely cited in coverage of the company’s expanded model‑building plans. (sec.gov)) NVIDIA concurrently shipped Nemotron 3 Super as a high‑parameter open‑weight release (reported at ~128 billion parameters) and has positioned its proprietary services and DGX Cloud training as the route for enterprises that want managed, production‑grade integrations. (decrypt.co))
