One model for the document workflow
A credit review can require extracting figures from tables, interpreting a chart, checking a linked web page and applying internal rules. NVIDIA says Super VL can handle those inputs, follow instructions and use tools. Companies can fine-tune it on their own data, tools and policies, then choose infrastructure according to their security and data-storage requirements.
The rest of the release covers local workloads:
Super VL is not NVIDIA’s first model for visual data. Nemotron Nano 2 VL already handled documents and images. In April, NVIDIA released the 30-billion-parameter Nemotron 3 Nano Omni, which works with text, images, audio and video. The new model’s place in the lineup is different: it brings document and image analysis into the larger Super tier, alongside models for agent reasoning and long-running workloads.
That could simplify some agent systems. An insurance workflow, for example, might need to read scanned documents, extract structured data, compare it with a policy and flag issues for an employee. Using one model for text and visual material could reduce the need to pass data between separate recognition systems. It does not remove the need to check whether the model extracted the facts correctly or applied the right rules.
Other open multimodal models include Meta Llama 4 and Google Gemma 4. NVIDIA’s pitch for Super VL rests on its larger Nemotron base, its tuning and deployment tools, and results from partners that adapted the model for narrower tasks.
Fine-tuning is the claim behind the scores
NVIDIA says dozens of organizations are fine-tuning Super VL. Named projects include Scale AI work for a Skydance financial agent handling supplier credit data and an EY agent for mergers-and-acquisitions reviews; AfterQuery’s contract analysis for Hebbia; AI21 and Alice projects in cybersecurity; and Mercor agents for coding tasks. Ramp, Dream, Lila Sciences and R1 are also tuning the model.
Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys are listed as platforms integrating Super VL into their products.
NVIDIA’s chart compares the base model with versions fine-tuned for five specialized tasks. It reports gains of 9–44%, and says each tuned version scored higher than the Claude or GPT model selected for that comparison.
Those are results for different models tuned to different tasks, not evidence that the untuned Super VL beats Claude Opus 5 or GPT-6.1 Sol overall. NVIDIA does not provide task descriptions, sample sizes or testing protocols alongside the chart. I think that omission matters more than the headline comparisons: buyers need to know whether the benchmarks resemble their own documents and failure cases before treating the gains as evidence for deployment.
Fast output is not a fivefold faster workflow
NVIDIA also says Super VL can generate output up to five times faster than a comparable open model. Its chart plots the Artificial Analysis Intelligence Index against output tokens per second with 100,000 input tokens. Super VL is shown at about 625 output tokens per second and an index score of about 26; the fivefold comparison starts at Inkling Small, at about 125 tokens per second and a similar score.
NVIDIA says Artificial Analysis measured the 625-token-per-second result on four Blackwell B200 GPUs. But the chart does not show Super VL leading on the index: it scores 26, compared with 42 for GLM-5.3-Flash and 40 for Qwen3.8-Flash-Next. The index combines ten evaluations across areas including professional work, coding, science and reasoning. Image processing is assessed separately, so the overall score does not capture all of Super VL’s document-analysis ability.
Fast generation can matter when an agent makes many calls during a long task. It is not the same as finishing the task five times faster: input processing, external tools, checking results and correcting errors all take time. The announcement gives a generation-speed comparison, not a measurement of end-to-end work.
Bolt takes the work off the cloud
Bolt 74B-A9B uses built-in compression to 2 bits, which NVIDIA says lets it fit in 24 GB of memory and run alongside other applications on RTX Spark and DGX Spark systems. Its intended jobs include web research and coding or debugging with local agent tools, without sending those tasks to a cloud model.
NVIDIA calls Bolt the best open model in its class on PinchBench and second among open models on BrowseComp. The announcement gives no scores and does not say which models are included in the comparison. Bolt is available on Hugging Face and supports llama.cpp, Ollama, LM Studio and vLLM. NVIDIA lists GeForce RTX computers, DGX Station and RTX PRO workstations among supported systems.
Bolt E4B is designed for 4 GB of memory, with voice assistants, interactive characters and embedded interfaces as target uses. NVIDIA says it supports multi-turn dialogue and can use tools while taking the application’s current state into account. On its planned October 23 release, it will target GeForce RTX, Jetson and RTX/DGX Spark.
Super VL’s open weights are available on Hugging Face; NVIDIA also names OpenRouter, ModelScope and NVIDIA NIM as launch access points. NVIDIA says the models use the OpenMDW license, which permits commercial use, fine-tuning and redistribution. It is also releasing training recipes and datasets, with tuning materials available through NeMo and a technical architecture guide.
Organizations can use NVIDIA’s hosted API with a request limit for free development. Partners offer hosted APIs at their own prices, including through OpenRouter. For production, NVIDIA lists AI Enterprise at $4,500 per GPU per year. That is a software-license price, not the total cost of GPUs, hosting and operating Super VL.
NVIDIA says Super VL is optimized for Blackwell and Hopper systems, including DGX Spark, DGX Station and RTX PRO 6000 Blackwell GPUs, and can also run on Ampere-generation GPUs. Its reported training scale includes about 26 trillion tokens across pretraining and continued pretraining, 3.8 million fine-tuning examples, and reinforcement learning in about 100 open environments.
The release follows NVIDIA’s agreement to buy Hugging Face for about $12.93 billion. The deal has not closed; NVIDIA expects to complete it in the first half of 2027, subject to regulatory approval and other conditions. The company has committed to keeping Hugging Face available to other model developers and chipmakers.
My guess is that the practical test is less whether Super VL can top a partner benchmark than whether a company can tune it to its own documents and procedures at a viable operating cost. NVIDIA has laid out the models, access routes and production-license price; the missing piece is evidence that those ingredients deliver the same gains inside a buyer’s workflow.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X