# Nvidia releases Nemotron 3.5 Super VL with open weights, training data and reinforcement-learning environments

The 120-billion-parameter mixture-of-experts model activates 12 billion parameters per token and handles a one-million-token context, and its reported wins over Claude Opus 5 come from task-specific post-training that no outside group has yet reproduced.

- Published: 2026-10-17T06:12:13.157Z
- Canonical: https://polylog.news/ai/2026-10-17/nvidia-releases-nemotron-3-5-super-vl-with-open-weights-trai
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [The Decoder](https://the-decoder.com/nvidias-nemotron-3-5-super-vl-is-an-open-model-built-to-run-enterprise-ai-agents-nonstop/), [VentureBeat](https://venturebeat.com/technology/nvidias-new-open-nemotron-3-5-super-vl-with-post-training-beats-claude-opus-5-gpt-6-1-sol-on-specialized-task-benchmarks), [MarkTechPost](https://www.marktechpost.com/2026/10/15/nvidia-nemotron-3-5-super-vl-release/), [Artificial Analysis](https://artificialanalysis.ai/articles/nemotron-3-5-super-and-bolt), [Serverflow](https://serverflow.ru/blog/novosti/nvidia-predstavila-nemotron-3-5-super-vl-otkrytaya-multimodalnaya-model-na-120-mlrd-parametrov-dlya-/), [OrcaRouter](https://www.orcarouter.ai/blog/nemotron-3-5-super-vl-launch)

Nvidia released Nemotron 3.5 Super VL on October 15. It is an open-weight multimodal model built for agents that run continuously inside enterprise systems, [according to The Decoder](https://the-decoder.com/nvidias-nemotron-3-5-super-vl-is-an-open-model-built-to-run-enterprise-ai-agents-nonstop/). The model is a mixture-of-experts network with 120 billion total parameters and 12 billion active parameters. It reads text, images and video clips of up to two minutes, [as the Russian-language channel AI ML Big Data summarized](https://t.me/ai_machinelearning_big_data/11145). The context window reaches one million tokens, and Nvidia licenses the model under OpenMDW-1.1.

This release includes more than most open-weight launches. Nvidia published a base checkpoint, a post-trained checkpoint and a 4-bit compressed checkpoint. It also released the datasets, the training recipes and the reinforcement-learning environments, and made the model available on Hugging Face, OpenRouter, ModelScope and its own NIM service, [The Decoder reported](https://the-decoder.com/nvidias-nemotron-3-5-super-vl-is-an-open-model-built-to-run-enterprise-ai-agents-nonstop/). [Launch coverage](https://www.marktechpost.com/2026/10/15/nvidia-nemotron-3-5-super-vl-release/) describes a hybrid architecture that combines Mamba-2, mixture-of-experts and attention layers, with a C-RADIO vision encoder and 512 routed experts per expert layer. The state-space layers make the million-token context feasible, because they limit growth in the key-value cache, the memory that stores earlier tokens during generation (also called the KV-cache), compared with what a pure-attention architecture of the same depth would require.

On serving speed, Nvidia says the model generates more than 600 tokens per second on a single GPU, up to five times as fast as open competitors in its class. It credits built-in multi-token prediction, a speculative-decoding method that speeds generation on DGX Spark by up to 2.3 times, and NVFP4 quantization with distillation that Nvidia says keeps BF16 accuracy, [according to Serverflow](https://serverflow.ru/blog/novosti/nvidia-predstavila-nemotron-3-5-super-vl-otkrytaya-multimodalnaya-model-na-120-mlrd-parametrov-dlya-/). These are vendor figures.

Independent verification is more limited. [Artificial Analysis](https://artificialanalysis.ai/articles/nemotron-3-5-super-and-bolt) scores the model at 26 on its Intelligence Index, well above the average for open-weight models of similar size. It also reports roughly twice the predecessor's index score with about 60% fewer output tokens per task. The stronger claim comes from [VentureBeat](https://venturebeat.com/technology/nvidias-new-open-nemotron-3-5-super-vl-with-post-training-beats-claude-opus-5-gpt-6-1-sol-on-specialized-task-benchmarks), which reports that after post-training, the model beats Anthropic's Claude Opus 5 and OpenAI's GPT-6.1 Sol on specialized task benchmarks. That comparison pits a domain-tuned model against general-purpose ones, and [OrcaRouter noted](https://www.orcarouter.ai/blog/nemotron-3-5-super-vl-launch) that no independent scoreboard existed at launch. The weights, the license and the architecture can be verified now. The head-to-head wins over closed frontier models are Nvidia's claim until an outside group reproduces them on tasks Nvidia did not select.

## What this means

The mechanism runs vertically. Nvidia gives away the model layer, including the reinforcement-learning environments enterprises usually lack, to raise demand for the hardware those recipes are tuned for (NVFP4 targets Nvidia's Blackwell-generation GPUs). Closed API vendors selling mid-tier multimodal agent capacity have the most to lose, because a company that can post-train a 12-billion-active model on its own workflows has less reason to pay frontier per-token rates for narrow tasks. Chinese open-weight labs also lose some ground with U.S. enterprises whose compliance teams avoid Chinese weights. The market already treats Nvidia's models as marketing for its chips. The consensus underrates the released environments and data, which turn the model into a fine-tuning starting point rather than a finished product. The precedent is Meta's Llama 2 release in 2023, which produced a large fine-tuning ecosystem and pushed prices down for GPT-3.5-class APIs within months. The likelier outcome is that independent evaluations over the coming weeks place Super VL below closed frontier models on general agentic suites, while domain-specific post-trained derivatives appear on Hugging Face within months, provided enterprise legal teams accept the OpenMDW-1.1 terms. The read is wrong if outside reproductions show the post-trained advantage disappears on held-out tasks, or if derivative counts stay far below those for Qwen and GLM bases.

## What to watch

- Whether any independent group reproduces the post-trained wins over Claude Opus 5 and GPT-6.1 Sol on tasks Nvidia did not choose, which would show whether the specialization claim holds outside the launch materials.
- The number of fine-tuned derivatives of Super VL on Hugging Face compared with Qwen and GLM bases, because derivatives measure real adoption better than download counts.
- Whether closed labs cut prices on their mid-tier multimodal models in the coming weeks, which would show they see Nvidia's open release as a pricing threat.