Nvidia's latest open-model release gives companies more choices about where their AI runs, from a data center they operate to a computer on an employee's desk. The practical distinction is how much useful capability fits into the memory they can afford.
On October 15, Nvidia released Nemotron 3.5 Super VL, a 120-billion-parameter model for agents that read documents, use tools and carry out multistep work. It also released Bolt 74B-A9B 24GB for local machines. A smaller Bolt E4B 4GB model is scheduled for October 23. These are releases with different hardware requirements, rather than one model that performs identically everywhere. Nvidia announcement
Downloading model weights, the numerical settings learned during training, lets a company operate the model without sending every request to its maker's hosted service. Somebody still has to supply computing equipment, maintain software and decide whether an agent is doing its assigned job correctly.
The documents accompanying this release explain how Nvidia is trying to make that calculation more attractive.
A 74-billion-parameter model in 24 gigabytes
Bolt begins with a larger model and compresses it. Nvidia's technical report describes reducing Super's 120 billion total parameters to 74 billion, then storing much of the resulting network at approximately two bits per parameter. More ordinary, higher-precision formats require substantially more memory. Bolt technical report
That process reduces the compressed model's weights from 148 gigabytes in BF16, a 16-bit format, to 22.1 gigabytes. The report calls the quantization lossy: reducing numerical detail changes the model. Additional training helps recover capabilities after compression, including its ability to interpret images and use tools.
The 24-gigabyte target covers text-only serving, excluding the vision encoder and projector. Nvidia designed for one request at a time and 128,000 tokens of context, the pieces of information a model processes. Longer inputs and simultaneous requests require more memory.
Those details help a technology buyer distinguish a downloadable file size from the requirements of an application employees can actually use. A model that fits by itself may leave inadequate space for other programs or for a busier workload.
Fast enough depends on the assignment
Nvidia's Bolt research page reports about 84 output tokens per second on a Surface Laptop Ultra with 128 gigabytes of memory, using its specified software and acceleration method. A token-generation rate describes how quickly text appears under those conditions. Bolt results
The same table contains a useful qualification. Two comparison models, Qwen3.6 35B-A3B and Gemma 4 26B-A4B, generate tokens faster in the displayed test. They also finish its agent replay in roughly 454 and 481 seconds, compared with about 601 seconds for Bolt. Nvidia presents Bolt as balancing intelligence and speed at a given memory size, rather than winning every speed measure.
An agent replay follows recorded sequences of model activity. It helps compare execution time under controlled conditions. A company choosing an assistant must also measure whether it completes the required assignment correctly. The fastest replay and the best accepted answer can belong to different configurations.
This makes task selection consequential. A document assistant, a coding agent and a conversational character place different demands on memory, response time and acceptable mistakes. Buyers need results for the relevant job, with the hardware and settings stated beside them.
Customization supplies a separate test
Scale AI, which sells model-customization services, reported early work with Super VL at Skydance and EY. At Skydance, it says an agent examines vendor credit documents and invoices to identify credits owed to the company. Scale reports matching the accuracy of the existing frontier model at about one-fortieth of its inference cost and five times the response speed. Scale's account
For EY, Scale describes answering acquisition questions across virtual data rooms. It reports matching the incumbent model's accuracy on held-out questions at about one-eighteenth of the inference cost. These are vendor-reported evaluations of specialized systems. They do not establish savings after training, data preparation and ongoing operations are included.
Francis deSouza, Scale's chief executive, said in the company's announcement that enterprises need to “prove which model is best for each job.” Scale's commercial interest is in conducting that customization, which gives readers a reason to examine the examples and the comparison methods together.
The narrower task and evaluation set matter. Matching accuracy on invoices or deal questions supports a decision about that application. It does not establish that the customized model matches a larger general-purpose model across unrelated work.
Serving more users changes the design
Nvidia's Super VL technical report explores a different operating problem: using server hardware to answer many requests while keeping responses fast enough for each person. It reports up to 4,856 output tokens per second per GB200 GPU while maintaining at least 50 tokens per second per user in the tested configuration. Super and Lightning report
Those figures describe aggregate throughput and individual response speed, respectively. A large combined number can coexist with slower responses for each person if too many requests share the equipment. Comparing a server figure with a single-user laptop figure without preserving that distinction would give a misleading impression.
The report separately gives Super a 51.78% result on SWE-bench Pro, a software-engineering test. That score concerns a named benchmark and configuration. Buyers comparing it with other coding models need results from that same test, rather than percentages from unrelated coding evaluations.
Nvidia's developer account also describes reusing processing from instructions and conversation history that agents send repeatedly. Avoiding repeated computation can improve efficiency without changing a user's question. Its serving tools coordinate where requests run and where their saved context remains available. Nvidia technical explanation
Open weights come with operating responsibilities
Super VL's model card lists commercial and noncommercial use and points to the OpenMDW 1.1 license. The license permits broad use of the supplied materials without a model-license charge. Distributors must preserve the agreement and applicable notices. It also leaves users responsible for other rights and permissions their use requires. Model card, license
For a company comparing hosted and local deployment, the next step is a priced trial on its own tasks: equipment, operation, customization, accepted answers and employee time. Nvidia has supplied more combinations to test. The scheduled smaller Bolt release on October 23 will add another memory target, making its actual requirements and measured results the next documents to examine.
Common questions
Does Bolt’s 24-gigabyte budget include vision?
It excludes the vision encoder and projector. Bolt report
Why does Super have both 120B and 12B figures?
It contains 120 billion parameters and activates about 12 billion per token. Its mixture-of-experts design selects parts of the network for each step. Technical guide
What can Super VL accept?
It accepts text, images and sampled video frames. Supply documents as page images with accompanying text. It generates text. Model card
How long can Super’s context be?
The model supports up to one million tokens. Prompts, images, history, tool results and generated output share that allowance. Deployment settings may use less. Model card
Which serving tools does Nvidia document?
Its model card provides configurations for vLLM, SGLang and TensorRT-LLM, with hardware-specific settings. These are serving recipes, not a promise of identical performance. Model card
Can companies adapt the model to their documents?
Nvidia describes supervised fine-tuning and low-rank adaptation through NeMo, plus reinforcement learning for agent tasks. Training examples and evaluation should match the intended workflow. Customization guide
What must be retained when distributing the model?
OpenMDW 1.1 requires a copy of the agreement and applicable copyright and origin notices when distributing covered model materials. OpenMDW 1.1
Does OpenMDW add restrictions to generated output?
The agreement imposes no restrictions or obligations on output use, modification or sharing. Separate rights belonging to other people can still require permission. OpenMDW 1.1
What training material did Nvidia release?
Its technical guide lists tool-use, conversation, finance, instruction-following, long-context and vision-language datasets. Developers must review each dataset’s own terms and suitability. Dataset guide
Why separate total throughput from an individual user’s speed?
Total throughput adds output across simultaneous requests. Per-user speed describes one stream. Nvidia’s report states both so its server results can be interpreted in context. Technical report
