Glonce

Nvidia releases open Nemotron 3.5 Super VL, a 120B model for enterprise agents

modelsThe Glonce editorsThe Decoder🔥68Glonce rating68Impact60Novelty60Relevance100Surprise35Players85Credibility75
Nvidia releases open Nemotron 3.5 Super VL, a 120B model for enterprise agents

Nvidia has released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters built for enterprise agents that run continuously in the background and process text, images, documents and video. It uses a Mixture-of-Experts architecture in which only 12 billion of the 120 billion parameters are active per token. Nvidia developed it with the Nemotron Coalition, a group that includes Mistral AI, Thinking Machines Lab, Perplexity and Black Forest Labs.

Two smaller Bolt models launch alongside it. Nemotron 3.5 Bolt 74B-A9B 24GB runs agent tasks on local machines; it is a hybrid Mixture-of-Experts model with 9 of its 74 billion parameters active per step, squeezed into 24 GB of memory with 2-bit quantization of the weights. The smaller Bolt E4B 4GB, aimed at voice assistants, interactive characters and device interfaces, has not shipped yet and is due later this month. Nvidia claims both offer the "highest intelligence per gigabyte of memory." Super VL and Bolt join the existing Nemotron lineup: Lightning, released in August, is built for high throughput, and Ultra handles hard reasoning tasks. Every variant is optimized for Nvidia hardware.

Nvidia distilled Super VL from Nemotron 3 Ultra, its largest reasoning model. On instruction following, tool calling, multi-step business processes and image analysis, Super VL is supposed to match Ultra's accuracy with about one-fifth the parameters. In agent tests from Hermes and LangChain it scores higher than Ultra, according to Nvidia, and it beats its predecessor Nemotron 3 Super on the coding benchmarks SWE-bench and Terminal-Bench. On a single GPU, Nvidia says Super VL delivers up to five times the throughput of similarly sized open models on agent tasks, citing the Artificial Analysis leaderboard. In an independent test by EXO Labs on the DGX Spark, no other model was both more accurate and faster.

Efficiency work is a large part of the pitch. Super VL is compressed with Nvidia's 4-bit NVFP4 format: on Blackwell GPUs it runs entirely in 4-bit precision, while older cards store only the weights at 4 bits and do the math at 16 bits. Nvidia trained the compressed version to match the uncompressed original's outputs. A built-in draft component proposes several tokens ahead for the main model to verify in one forward pass; on the DGX Spark, Nvidia recommends its own LiLiCorr method, which it says speeds up text generation by up to 2.3x versus running without speculative drafts. Because agents resend the same instructions, tool descriptions and conversation history with each request, the model caches those computations, and Nvidia's Dynamo software routes requests from one session to the GPU that holds its cache. On 16 B200 GPUs, Nvidia says this serves eight times as many concurrent users and cuts time to first token by 85 percent.

On vision, Nvidia targeted OCR, tables, charts, multi-page documents and video. Super VL can read a chart, link it to a table elsewhere in the same document and summarize both without a separate image model. On the CharXiv chart benchmark it scores 80.8 percent according to Nvidia, against 63 percent for the older Nemotron 3 Nano Omni; with Python code execution the score rises to 84 percent. Nvidia says it ranks among the best open models on the OCRBench v2 leaderboard.

The Bolt 74B-A9B model handles everyday local agent work such as multi-step tasks, tool calls, web search, writing and debugging code, or working out which PC components fit a system. It is optimized for RTX Spark, a new chip for AI-capable Windows laptops, and for the DGX Spark workstation with up to 128 GB of shared memory. It also runs on GeForce RTX PCs, DGX Station and RTX PRO workstations, and supports llama.cpp, Ollama, LM Studio and vLLM. The article notes that 24 GB is a high bar: only pricey high-end hardware has that much memory, so the local agent is for now aimed at developers and owners of powerful workstations.

Nvidia says dozens of organizations are already fine-tuning Super VL, mainly in industries with heavy document loads and sensitive data. Scale AI adapted it for an agent handling vendor credits at media company Skydance, with costs expected to fall to one-fortieth and response time to one-fifth. Ramp uses it to turn receipt photos into structured data and expects ten times faster responses. EY uses it for acquisition due diligence and Hebbia for analyzing credit agreements. In cybersecurity, AI21 fine-tuned it for security firm Alice, with accuracy improving by 40 percent according to Nvidia; Dream uses it to hunt for vulnerabilities and CrowdStrike is building agents to defend enterprise networks. Salesforce plans to automate government workflows, R1 is auditing medical billing, and Perplexity, Dassault Systèmes and Mercor are also using it. Against closed frontier models, Nvidia pitches Super VL to companies that want to fine-tune on their own data and run it on their own servers, free from a provider's terms of service.

The release is more open than many open-weight models, which typically share only weights. Nvidia is publishing a base version, a post-trained version and a 4-bit compressed version, along with datasets, training recipes and reinforcement learning environments. Nvidia says the model was trained on over 25 trillion tokens in pre-training and continued pre-training, with post-training and reinforcement learning adding roughly four million examples across about 100 open training environments. Datasets on Hugging Face cover agent tool use, instruction following, long texts and image understanding, plus finance, safety and fairness. One, built with the Indian Institute of Technology Madras and its AI lab AI4Bharat, holds about 2.83 trillion tokens of text in 22 Indian languages. For customization Nvidia provides its NeMo framework, Python pipelines and open datasets. Enterprises can also get the model through Nvidia NIM as a ready-to-deploy microservice with validated configs and support, updated every two weeks. Weights for Nemotron 3.5 Super VL 120B-A12B and Nemotron 3.5 Bolt 74B-A9B 24GB are on Hugging Face.

The article's author calls the strategy self-serving: the more companies run their own open models, the more compute hardware they need, and open models will likely stay optimized for Nvidia for the foreseeable future, while Meta, Google, Anthropic and OpenAI build their own chips and tune models for them. The author notes that Nvidia recently invested $12.9 billion in Hugging Face.

Key facts

  • Nemotron 3.5 Super VL is an open multimodal Mixture-of-Experts model with 120 billion parameters, of which 12 billion are active per token, built for enterprise agents.
  • Nvidia says the model, distilled from Nemotron 3 Ultra, matches Ultra's accuracy on agent and image tasks with about one-fifth the parameters, and delivers up to five times the throughput of similarly sized open models.
  • Bolt 74B-A9B 24GB (9 of 74 billion parameters active, 2-bit weights) targets local machines; Bolt E4B 4GB for voice assistants is due later this month.
  • Weights come in base, post-trained and 4-bit versions, with datasets, training recipes and reinforcement learning environments; weights are on Hugging Face.
  • Nvidia names dozens of organizations fine-tuning the model, including Scale AI, Ramp, EY, Hebbia, AI21, CrowdStrike and Salesforce, though customer savings are expectations.

Why it matters

Super VL pairs an open model with unusually open ingredients: datasets, training recipes and reinforcement learning environments ship with the weights, which the article says makes it more open than many open-weight models that share only weights. Nvidia's pitch is a distilled model that it claims matches its largest reasoning model, Ultra, on agent and image tasks at about one-fifth the parameters, aimed at long-running agents that process large task volumes over days or weeks. The article's author adds that the openness serves Nvidia: everything is tuned for its hardware, so getting the most from Nemotron means buying Nvidia systems.

Who it affects

Enterprises that want to fine-tune a model on their own data and run it on their own servers, free from a provider's terms of service, are the main target; Nvidia cites an insurer screening claims for fraud and a financial firm scoring creditworthiness by its own criteria. Early users named by Nvidia are in finance, document-heavy work and cybersecurity. Developers and owners of powerful workstations are the audience for the local Bolt agent, since 24 GB of memory is a high bar for most consumers. Open-model developers also get the datasets, including a 22-language Indian text collection.

How to use it

Weights for Nemotron 3.5 Super VL 120B-A12B and Nemotron 3.5 Bolt 74B-A9B 24GB are on Hugging Face. Super VL runs on DGX Spark, DGX Station, RTX PRO and RTX desktops, on Jetson modules at the edge, and in data centers and the cloud. Enterprises can take it through Nvidia NIM as a microservice with validated configs and support, updated every two weeks. Nvidia provides its NeMo framework, Python pipelines and open datasets for customization. Bolt supports llama.cpp, Ollama, LM Studio and vLLM and is optimized for RTX Spark and DGX Spark. Bolt E4B 4GB is not out yet. The source states no price or licence name for the weights.

How solid is it

Treat the performance numbers as vendor claims. The accuracy match with Ultra, the higher scores in Hermes and LangChain agent tests, the up to five times throughput, the 80.8 percent CharXiv score, the up to 2.3x LiLiCorr speedup, the eight times concurrent users and 85 percent faster first token on 16 B200 GPUs, and the 40 percent accuracy gain at Alice are all Nvidia's figures. The one independent test mentioned is by EXO Labs on the DGX Spark, where no other model was both more accurate and faster; the throughput claim cites the Artificial Analysis leaderboard. The Skydance and Ramp savings and speedups are expectations, not measured results. The article gives no numeric SWE-bench or Terminal-Bench scores.

Risks and caveats

Hardware lock-in is the central caveat: all variants are optimized for Nvidia, and the article's author expects open models to stay that way. The local Bolt model needs aggressive 2-bit quantization to fit in 24 GB, and the source does not say how its quality compares to the full-precision model. At current GPU prices most consumers cannot justify hardware with that much memory. The smaller 4 GB model has not shipped. Customer deployments are described by Nvidia, and some, such as Salesforce's government workflows, are plans rather than live use.

“highest intelligence per gigabyte of memory”

— Nvidia's claim about both Bolt models, as reported by The Decoder