Nvidia released two open-weight agent models on 15 October 2026 and said a third is coming on 23 October. Nemotron 3.5 Super VL 120B-A12B and Nemotron 3.5 Bolt 74B-A9B 24GB are available now; Nemotron 3.5 Bolt E4B 4GB is announced, not shipping (Nvidia’s announcement post). TechCrunch reported the launch on 19 October and called it a Thursday launch, which matches Nvidia’s own dated post and both model cards (TechCrunch).
The parameter counts are not the interesting part. The question worth answering is what “open” buys a team that wants to post-train an agent on its own data, and where the released material stops short of the marketing around it.
What each model is for and where it runs
Super VL is the enterprise system: 120 billion parameters with 12 billion active, built for multimodal reasoning and specialised agents, with base, post-trained and NVFP4-quantised checkpoints plus training recipes, data and reinforcement-learning environments (Nvidia developer blog). It is distributed through Hugging Face, OpenRouter, ModelScope and NVIDIA NIM.
Nemotron 3.5 Bolt 74B-A9B 24GB is the local model: 74 billion parameters with 9 billion active and a 256K context window, shipped as a 24GB quantised checkpoint with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM. Nvidia says it is built for RTX Spark laptops arriving on 16 October, and for DGX Spark workstations. Bolt E4B 4GB, due 23 October, targets conversational and physical agents on GeForce RTX, Jetson and RTX or DGX Spark hardware. Nvidia’s partner lists include Dell under hardware and Microsoft Foundry under cloud platforms; TechCrunch names Microsoft and Dell among the partners releasing the new AI PCs.
What the OpenMDW-1.1 licence permits
Both Hugging Face cards name the OpenMDW-1.1 licence and state that the models are “ready for commercial or non-commercial use” (Super VL card; Bolt card). Reading the licence text itself (openmdw.ai), the grant is broad: permission to deal in the model materials “without restriction”, covering use, copying, modification and distribution, with no field-of-use, royalty or geographic limits, no acceptable-use or prohibited-use list, and no copyleft terms.
Three things are required or changed. You keep a copy of the licence and the notices that apply to whatever you redistribute. Your rights terminate if you file or maintain a patent or copyright suit over the model materials, unless you are responding to a suit brought against you first. The licensee handles third-party rights and consents. Outputs sit outside the licence: it imposes no conditions on how you use or share what the model generates.
One correction to the usual shorthand. OpenMDW-1.1 is a permissive Linux Foundation licence, and its stewards describe it as intended to conform to the Open Source Definition. It is not yet on the SPDX licence list, where LicenseRef-OpenMDW-1.1 is the recommended identifier. The models are open-weight: the downloadable files are governed by this licence, while hosted and NIM or enterprise access carries Nvidia’s separate agreements, whose operative text we could not read from the page Nvidia links (it returned only a navigation shell when checked on 19 October).
Where “open” stops short
The stewards’ FAQ carries a caveat that matters more than any single clause: the licence “does not mandate completeness”. It governs whatever a provider chooses to release under it.
Super VL’s card shows how that plays out. Alongside more than 100 public datasets, it lists categories the release leaves out: private non-publicly accessible third-party datasets including Global Regulation, a TAUS translation memory, Scale HLE, HackerRank coding data and search RL data generated with Gemini 3 and GPT-5, plus Nvidia’s own private datasets. The weights are open and the training mix is not, so anyone post-training on their own data supplies the replacement.
What the 5x speed figure actually measures
Nvidia’s headline is that Super VL is up to 5x faster than other models in its class. The precise version sits in Nvidia’s developer blog: on a single GPU, the model delivers up to 5x higher throughput for agentic tasks compared with other open models in its size category, measured on Artificial Analysis’s agentic tasks on a Baseten endpoint, at over 600 tokens per second. A separate vendor figure claims 4x faster task completion from higher throughput and token efficiency. The comparison set is open models, not the frontier labs, and the numbers are vendor-claimed. Nvidia also points to an independent DGX Spark evaluation by EXO Labs and to third-party leaderboards (Artificial Analysis, PinchBench, WebArena, BrowseComp, OCRBench v2), though it chose the comparisons and the configurations. TechCrunch attributes faster-than-Anthropic-and-OpenAI comparisons, after task-specific post-training, to Nvidia; the Nvidia pages cited here do not name those models.
Nvidia’s own table makes the intelligence-per-gigabyte case
The strongest support for what Bolt buys is Nvidia’s benchmark table for the 74B model (technical report). On the same RTX Spark hardware with llama.cpp, Bolt 74B-A9B records 83.99 tokens per second against 96.63 for Qwen3.6 35B-A3B and 100.56 for Gemma 4 26B-A4B. It is also slower end to end on Artificial Analysis’s local agent replay benchmark, at 600.8 seconds against 453.97 and 480.68.
Bolt still wins on most of Nvidia’s accuracy benchmarks, and the pitch is that a stronger model fits into 24GB rather than that it generates tokens faster. Two details qualify the comparison, both from Nvidia’s own pages. The technical report says the competing models were chosen because 4-bit builds fit inside 24GB, and its speed table scores them at 4-bit while Bolt runs at 2-bit with a trained speculative drafter, a decoding accelerator designed to preserve the model’s output distribution. The card’s accuracy table, by contrast, lists the competitors’ scores at BF16. And the “#1 in class” ranking depends on restricting the field: on PinchBench, the card’s own table puts Qwen3.6 35B-A3B ahead of Bolt, 88.07 against 87.82.
What self-hosting and post-training demand
Nvidia supports customisation through NeMo: AutoModel for supervised fine-tuning and LoRA, NeMo Data Designer recipes, NeMo Gym environments and open datasets, with a GitHub cookbook. Bolt ships an unquantised BF16 checkpoint for further customisation.
The hardware tiers differ sharply. On a DGX Spark workstation the published recipes keep concurrency to one or two sequences: the vLLM and SGLang configurations serve one concurrent sequence, and the TensorRT-LLM configuration caps batch size at two and context at 128K. The card lists 1M-token context for DGX Spark at that concurrency; the configurations built to serve many requests at once are the data-centre ones, 2x H100, 1x H200 and 1x GB200. Bolt’s 24GB figure is a 2-bit quantisation of a 74B model, supported only on Blackwell SM120 and SM121. Local deployment buys privacy and fit at the cost of serving only one or two sequences at a time, a narrow supported matrix and a compressed model.
Nvidia routes post-training work through partners including Scale AI, Mercor, Baseten, Databricks, Fireworks, Prime Intellect, Thinking Machines and AI21. No figure for the compute, cost or staffing of a post-training run on any of these models appears in the accessible sources.
What the named customers have confirmed
Nvidia says dozens of organisations are post-training Super VL on their own data, tools, policies and workflows. The companies it names are described as post-training, customising or integrating the model. None is described as running it in production.
Kari Briski, Nvidia’s vice president of open models and software, writes in the 15 October post that organizations can “post-train Super VL on their own data, deploy it on their own infrastructure and oversee what they build”. TechCrunch prints the same sentence with “own” in place of “oversee”; the wording quoted here is the post’s.
The numbers attached to those names carry different weights. Scale AI says inference cost for Skydance’s vendor-credit finance agent will fall by 40x and latency by 5x, a projection. Ramp expects up to 50% higher throughput and 10x lower latency, also a projection. AI21 and Alice report a 41% accuracy improvement for cyber threat intelligence work, a partner-reported result.
Where the frontier labs fit
TechCrunch notes that the labs which buy and invest in Nvidia silicon are also expanding to AWS Trainium, Cerebras and in-house chips. That is reporting and analysis rather than an Nvidia disclosure. Nvidia’s own partner lists include Amazon SageMaker JumpStart, Google Cloud’s Gemini Enterprise Agent Platform, Microsoft Foundry and Oracle Cloud Infrastructure, so the release is not confined to Nvidia-hosted deployment.
Where that leaves a buyer
For a team with a workload that fits inside 24GB and a reason to keep data off someone else’s servers, the licence is the strongest part of this release: permissive, commercially usable, with no acceptable-use list and no claim on outputs. For work that needs frontier-class reasoning, multi-user throughput or a documented route to a reproducible post-training result, the published material does not yet answer those needs, and the numbers on offer are Nvidia’s own comparisons, plus leaderboard placings and one partner-run evaluation that it cites.
