Nvidia Releases Nemotron 3.5 Super VL to Widen the Open-Weight AI Field

Nvidia released two open-weight multimodal models on Hugging Face on October 13: Nemotron 3.5 Super VL, a vision-language model with roughly 121 billion total parameters and about 13 billion active per token across 512 routable experts, and the smaller Nemotron 3.5 Bolt 74B-A9B. The releases follow an SEC filing in which Nvidia committed $26 billion over five years to open-weight AI models, a push aimed at closing a gap with Chinese labs whose open models grew from about 1.2% of global open-model usage in late 2024 to roughly 30% by the end of 2025, after Nvidia's Nemotron 3 Ultra scored 38 on the Artificial Analysis Intelligence Index versus 54 for Kimi K2.6.

Nvidia put a new multimodal model stack on Hugging Face this week, free for anyone to download, while the chip stocks that fund it were getting hammered on the trading floor. Nvidia updated Nemotron 3.5 Super VL and a smaller sibling, Nemotron 3.5 Bolt 74B-A9B, on Hugging Face on October 13. Both are open-weight releases, and they landed in the middle of a brutal week for the company's own hardware customers. Super VL is a vision-language model built on a hybrid Nemotron-H decoder, alternating Mamba and mixture-of-experts blocks. It runs roughly 121 billion total parameters but only about 13 billion active per token, spread across 512 routable experts, 22 of which fire for any given input. A separate 0.65-billion-parameter RADIO vision tower and projector handle images, reusing the image and video stack Nvidia built for its earlier Nemotron Omni models. Bolt 74B-A9B strips that down into a leaner, cheaper-to-run variant: aimed at developers who don't need the full multimodal package. Both models are free to download, modify, and run locally, and the developer community moved fast. The underlying code for Super VL had already been merged into Nvidia's own NeMo-RL repository on GitHub back in September. That support included an image-based reinforcement learning recipe running on stock vLLM. And llama.cpp contributors, who'd already built GGUF conversion paths for earlier Nemotron 3 and 3.5 releases, were positioned to extend that work to the new models almost immediately. That's the open-weight playbook: ship the weights, let Hugging Face and the GGUF crowd do the packaging, and your model ends up running on a hobbyist's gaming PC within days instead of waiting on anyone's enterprise API. Nvidia isn't doing this out of charity. According to a financial filing with the US Securities and Exchange Commission, the company committed $26 billion over five years specifically to build out open-weight AI models. That's a chip maker spending real money to make sure open models keep getting better. Not because Nvidia wants to compete head-on with OpenAI or Anthropic's closed frontier systems, but because every open model that runs well needs GPUs under it. A thriving open-weight ecosystem is demand generation with extra steps. It's also a response to a scoreboard Nvidia doesn't currently lead. When the company released Nemotron 3 Ultra on June 4, 2026, as its strongest open-weight reasoning model yet, it scored 38 points on the Artificial Analysis Intelligence Index. China's Kimi K2.6, by comparison, scored 54 on the same index, roughly in line with other releases from DeepSeek, Qwen, GLM, and MiniMax. Chinese labs now dominate the open-weight leaderboard across the board, and the shift happened fast: Chinese open models went from about 1.2% of global open-model usage in late 2024 to roughly 30% by the end of 2025. Nvidia doesn't have a model that beats that pack outright. What it has is a willingness to keep releasing, often, and across a range of sizes, so American developers have a steady stream of alternatives that don't require sending data to a lab in Beijing or Hangzhou. Nvidia Open Sources a Voice AI That Talks and Listens at Once https://startupfortune.com/nvidia-open-sources-a-voice-ai-that-talks-and-listens-at-once/ PersonaPlex-7B runs full-duplex, listening and speaking on the same model at once, built on Kyutai's open Moshi architecture. Nvidia also reports a time-to-first-token of roughly 170 milliseconds, on top of its interruption and turn-taking numbers. - open source voice AI that listens simultaneously https://startupfortune.com/nvidia-open-sources-a-voice-ai-that-talks-and-listens-at-once/ - Nvidia PersonaPlex enables natural overlapping conversation AI https://startupfortune.com/nvidia-open-sources-a-voice-ai-that-talks-and-listens-at-once/ That's the real function of Nemotron 3.5 Super VL and Bolt 74B-A9B. Neither is positioned as a frontier-beating release. They're infrastructure plays: models sized and architected so a startup can run vision-language tasks, document understanding, or agentic workflows on hardware it already owns. No subscription to a closed API, none of the latency or privacy tradeoffs that come with sending every request to a third party. For a developer deciding between a closed frontier model and an open one, that calculus has shifted meaningfully over the past year. A closed model like Claude or GPT-5 still wins on raw capability for the hardest reasoning tasks. But for a huge share of real production work, that's not the job. Document parsing, image classification, structured extraction: an open model that runs on your own GPUs at a fixed cost is the more rational choice. Nvidia is betting there's enough of that work to justify spending billions making sure the options keep getting better. The timing is almost accidental but still telling. The same week Nvidia pushed these weights out, Samsung and SK Hynix shares were sliding on fears that the AI memory boom has overheated, and South Korea's Kospi briefly halted program sell orders in a sidecar trigger. Hardware is volatile right now. Software that runs on hardware people already own looks a lot more durable by comparison, and Nvidia knows it. Also read: CoreWeave default odds hit 60 percent as credit swaps top 1,000 basis points https://startupfortune.com/coreweave-default-odds-hit-60-percent-as-credit-swaps-top-1000-basis-points/ • Gemini 4 Argon ties for third place on an AI benchmark Google helped build https://startupfortune.com/gemini-4-argon-ties-for-third-place-on-an-ai-benchmark-google-helped-build/ • Banks that financed the AI data center boom now want out of that debt https://startupfortune.com/banks-that-financed-the-ai-data-center-boom-now-want-out-of-that-debt/ This article is posted in AI News https://startupfortune.com/category/ai/ , check it out for more related stories. Nvidia's $12.9 Billion Bid for Hugging Face Signals a New AI Land Grab https://startupfortune.com/nvidias-129-billion-bid-for-hugging-face-signals-a-new-ai-land-grab/ Nvidia is reportedly nearing a $12.9 billion deal for Hugging Face, the clearest sign yet that open-weight AI labs have become Silicon Valley's hottest acquisition targets. The rush comes as Chinese open models like Qwen and Kimi K3 pull ahead of Meta's Llama in developer adoption, pushing American tech giants to buy their way into the open... - nvidia's acquisition of hugging face for billions https://startupfortune.com/nvidias-129-billion-bid-for-hugging-face-signals-a-new-ai-land-grab/ - open weight ai models becoming valuable acquisition targets https://startupfortune.com/nvidias-129-billion-bid-for-hugging-face-signals-a-new-ai-land-grab/ Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.