NVIDIA shipped a new open-weight flagship model on October 15, 2026: Nemotron 3.5 Super VL, a 120-billion-parameter mixture-of-experts system built for multimodal, always-on AI agents. The release lands in the middle of the busiest month for frontier model launches this year, arriving days after Google’s public rollout of Gemini 4 Argon and the same week OpenAI reportedly shelved its Astra 6.1 model over internal safety concerns. For developers trying to decide which open model to build on, Nemotron 3.5 Super VL raises an immediate question: does “open” still mean “behind,” or has NVIDIA closed the gap with the closed-weight leaders at the top of the leaderboards?
NVIDIA’s own blog network, developer.nvidia.com and blogs.nvidia.com, confirmed the launch alongside companion coverage from ZDNet and VentureBeat. The model is part of what NVIDIA calls the Nemotron Coalition, a collaborative open-model effort the company has been building out across 2026. Unlike a routine point release, Super VL ships with a specific pitch: agentic reasoning across PDFs, charts, tables, images, and webpages, wrapped in a license that reportedly permits commercial use, fine-tuning, and redistribution. That combination, open weights plus agentic multimodal reasoning, is what separates this release from the dozens of smaller open-source updates that cross Hugging Face every week.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Inside the 120B Parameter, 12B Active MoE Architecture
Nemotron 3.5 Super VL is a mixture-of-experts (MoE) model, meaning it does not activate all of its parameters for every token it processes. NVIDIA’s marketing material describes it as a 120-billion-parameter model with roughly 12 billion active parameters per token, which is where the “A12B” suffix in its full name comes from. NVIDIA’s technical documentation, however, gives a more granular figure: 121 billion total parameters spread across 512 routed experts, with 22 experts active per token. Add the multi-token-prediction head NVIDIA bundles with the model, and the documented total climbs to 124 billion parameters.
That distinction matters for anyone budgeting GPU memory. A dense 120-billion-parameter model would require far more active compute per forward pass than an MoE model that only lights up roughly a tenth of its weights at a time. It’s the same tradeoff that has driven most of the major 2026 open-weight releases, including DeepSeek V4.1 Flash and Liquid AI’s new LFM2.5-24B-A2B, both of which lean on sparse activation to keep inference costs down while still claiming large total parameter counts.
Licensing: NVIDIA’s OpenMDW Terms
According to reporting reviewed around the launch, Nemotron 3.5 Super VL is distributed under NVIDIA’s OpenMDW license, which is described as permitting commercial use, fine-tuning, and redistribution. That positions it closer to permissive open-source terms than the restrictive research-only licenses some labs still attach to their largest models. NVIDIA has not published a full breakdown clarifying whether every dataset bundled with the release carries identical terms, and that detail is worth checking directly in the model repository before building a production pipeline around it. What is confirmed is that NVIDIA paired the model release with open datasets on Hugging Face covering agentic tool-use and conversational data, along with finance-reasoning data, extending the open-source footprint beyond just the model weights themselves.
Built for Agents: Multimodal Tasks From PDFs to Tool Use
NVIDIA is not positioning Super VL as a general chatbot competitor to ChatGPT or Gemini. The company’s own framing, repeated across its developer blog and newsroom posts, is that the model targets “specialized, always-on agents” that organizations can post-train on their own data, tools, and policies. The model accepts three input types (text, images, and video), and the listed use cases include parsing PDFs, interpreting charts and tables, reading webpages and images, following multistep instructions, calling external tools, and handling coding tasks inside longer agentic workflows. That is a deliberately narrower pitch than a frontier assistant model, and it reflects where NVIDIA believes the commercial demand for open models actually sits in late 2026: inside enterprise agent deployments rather than consumer chat apps.
It is also a direct response to how the open-model conversation has shifted this year. Model families like GLM 5.3 Prime and Qwen3.8 Max have pushed hard on general-purpose reasoning benchmarks, while NVIDIA is betting that enterprises care more about task completion speed and deployability on their own infrastructure, a point relevant to the data-residency and security requirements VentureBeat reported NVIDIA emphasizing around this launch.
The Performance Claims NVIDIA Is Making, and What’s Verified
NVIDIA claims Nemotron 3.5 Super VL delivers “leading accuracy” on multimodal agentic tasks and up to 5x higher throughput than comparable models on the Artificial Analysis Intelligence Index, naming InclusionAI’s Ling 3.0 Flash and DeepSeek V4.1 Flash as the lower-throughput comparison points. VentureBeat’s coverage goes further, reporting NVIDIA’s claim that a post-trained version of Super VL beats Claude Opus 5.5 and GPT-6.1 Sol on specialized task benchmarks.
That claim deserves a caveat that is easy to lose in a headline: it is a company-reported result on NVIDIA’s own selected benchmark set, not an independent, reproducible leaderboard score verified by a third party. Public benchmark trackers that aggregate scores across dozens of evaluators, the kind referenced later in this piece, still place Claude Opus 5.5 and Gemini 4 Argon ahead of open models on general-purpose reasoning indices as of October 15. NVIDIA’s claim is narrower and task-specific: post-trained Super VL outperforming larger closed models on tasks it was specifically adapted for, not on broad general intelligence.
Long-Context and Coding Results on NVIDIA’s Research Page
NVIDIA’s research page for the family describes Nemotron 3.5 Super and Nemotron 3.5 Lightning as open and efficient Mixture-of-Experts Hybrid Mamba-Transformer models designed for reasoning, long-context understanding, and agentic workloads. Long context is the most relevant point for the document-heavy tasks Super VL targets. According to the page, Super’s 256K mid-training stage reduced its RULER-1M score, and a final 9-billion-token recovery stage at 1-million-token context restores the score from 33.5 to 81.0.
On coding and agents, the page states that Nemotron 3.5 Super achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra while using approximately 4.6× fewer active parameters. It also reports that agentic mid-training improves DeepSWE pass@1 from 6.4% to 10.4%, and that parallel agent execution followed by self-verification improves Terminal-Bench 2.1 from 62.7% to 71.9%, DeepSWE from 24.6% to 34.5%, and BrowseComp from 71.5% to 79.8% on the evaluated setups.
For serving, the page says MTP boosting improves average speculative-decoding acceptance length from 4.287 to 4.718 under greedy decoding and from 3.787 to 4.612 under probabilistic draft sampling. These are NVIDIA-reported results on the evaluated setups, so they carry the same caveat as the throughput claim, and teams should rerun them on their own workloads. Because the page covers Super and Lightning together, check which model a given score belongs to before comparing it against Super VL.
Mindbeam’s Post-Training Results: 72.7% to 83.9%
One concrete data point backs up NVIDIA’s post-training pitch. In a joint announcement distributed via PR Newswire, NVIDIA and a company called Mindbeam detailed a post-training run on Nemotron 3.5 Super VL (120B-A12B) aimed at a specific evaluation suite. Before adaptation, the model passed 72.7% of problems in what the companies describe as a Level 1 evaluation testing individual operations. After post-training, that pass rate rose to 83.9%, an 11.2 percentage point gain. It is a single case study rather than a broad benchmark, but it is the clearest externally reported evidence that NVIDIA’s “post-train it on your own workflow” pitch produces measurable gains rather than just marketing language.
Where to Get It: Hugging Face, NIM, OpenRouter, ModelScope
Open weights for Nemotron 3.5 Super VL are available through NVIDIA’s organization page on Hugging Face. NVIDIA also lists access through its own NIM microservices, OpenRouter, ModelScope, and the build.nvidia.com developer portal, where the model currently appears tagged as an early-access reasoning model for agentic, coding, tool-calling, and multimodal workloads. For teams that want managed deployment rather than self-hosted inference, NVIDIA AI Enterprise reportedly carries a production license priced at $4,500 per GPU per year, according to VentureBeat’s reporting. That figure applies to the enterprise software license layer, not a per-token API price, and no verified per-token API pricing for Nemotron 3.5 Super VL was available in the sources reviewed for this piece.
A typical self-hosted pull of the model through Hugging Face’s transformers library would look like a standard open-weight load:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "nvidia/Nemotron-3.5-Super-VL-120B-A12B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")NVIDIA documents context lengths up to 1 million tokens for Super VL, which is a different figure from Liquid AI’s separately announced 131,000-token window, a window that belongs to a different model.
The Nemotron 3.5 Bolt Edge Models
Super VL was not the only release under the Nemotron 3.5 banner. Coverage from wccftech describes NVIDIA introducing two smaller models under a “Nemotron 3.5 Bolt” umbrella, sized for roughly 4GB and 24GB memory footprints and aimed at edge and desktop deployment rather than data-center inference. The 24GB model, Nemotron 3.5 Bolt 74B-A9B, has 74 billion total parameters and 9 billion active parameters, uses native 2-bit quantization, and NVIDIA says it was compressed from the 120B-A12B Super VL model. The 4GB Nemotron 3.5 Bolt E4B followed, and NVIDIA released it on October 23, 2026. NVIDIA’s own framing calls this “open from data center to desktop,” pairing the 120-billion-parameter Super VL for server-side agents with the Bolt models for on-device or resource-constrained use cases. That spread, from a 4GB edge model to a 120-billion-parameter MoE flagship, mirrors Liquid AI’s own strategy with LFM2.5, which NVIDIA’s release arrived alongside in the same reporting window.
How Nemotron 3.5 Super VL Stacks Up Against Rival Open Models
The clearest way to read this release is against the other open-weight models that shipped in the same two-week window. The table below lays out the specs that are publicly confirmed as of October 15, 2026, and flags the gaps where companies have not disclosed numbers.
| Model | Total Params | Active Params | Context Window | License / Access |
|---|---|---|---|---|
| NVIDIA Nemotron 3.5 Super VL | 120B (121B per docs) | ~12B (22 of 512 experts) | Up to 1,000,000 tokens | OpenMDW (open weights); Hugging Face, NIM, OpenRouter, ModelScope |
| Liquid AI LFM2.5-24B-A2B | 24B | 2B | 131,000 tokens | Open weights via Liquid Docs |
| DeepSeek V4.1 Flash | Not disclosed in this comparison | Not disclosed | Up to 1,000,000 tokens (per API setup) | Open weights; API access |
| InclusionAI Ling 3.0 Flash | Not disclosed | Not disclosed | Not disclosed | Named by NVIDIA as throughput comparison |
Liquid AI’s entry is the most fully specified of the group: a 40-layer MoE design combining 30 gated short-convolution layers with 10 grouped-query-attention layers, using 64 experts per MoE block with top-4 routing. It is explicitly built for on-device reasoning and tool-calling rather than data-center-scale agentic workloads, which is why its total parameter count sits so far below Super VL’s. NVIDIA, by contrast, documents context lengths up to 1 million tokens for Super VL, a figure that matters for a model pitched at document- and chart-heavy multimodal tasks where context length is often the limiting factor.
The Closed-Model Benchmark Gap
Open-weight releases do not exist in a vacuum, and the closed-model leaderboard moved just as fast this week. BenchLeader’s October 15 ranking, which reconciles 218 benchmarks from 42 publishers into a daily composite score, places Claude Opus 5.5 first at 71.6, Gemini 4 Argon second at 71.1, and Claude Fable 5.1 third at 69.9. Separately, the BenchLM tracker, covering more than 221 models across 629 benchmarks, reports that Claude Opus 5.5 and GPT-6.1 Sol tie under its 90% statistical interval rule, with GPT-6.1 Sol retaining 97% of the top score at a 50% lower output price.
| Model | Type | BenchLeader Score (Oct 15) | API Pricing (per million tokens) |
|---|---|---|---|
| Claude Opus 5.5 | Closed | 71.6 (rank 1) | Not covered in this comparison |
| Gemini 4 Argon | Closed | 71.1 (rank 2) | $2.00 input / $10.00 output |
| Claude Fable 5.1 | Closed | 69.9 (rank 3) | Not covered in this comparison |
| GPT-6.1 Sol | Closed | Ties Claude Opus 5.5 under 90% interval (BenchLM) | 50% lower output price than top score, per BenchLM |
| NVIDIA Nemotron 3.5 Super VL | Open | Not yet on general leaderboards; NVIDIA claims task-specific wins over Claude Opus 5.5 and GPT-6.1 Sol | Open weights free; NVIDIA AI Enterprise license $4,500/GPU/year |
Read together, the two tables tell a consistent story. On broad, general-purpose reasoning, the closed frontier models still hold the top three leaderboard slots, with less than two points separating first and third place. NVIDIA is not trying to contest that leaderboard directly with Super VL. Its claimed wins are scoped to specialized, post-trained tasks, which is a narrower and more defensible claim than “beats GPT-6.1 Sol,” even if some headlines will compress it to exactly that.
Why NVIDIA Keeps Betting on Open Weights
NVIDIA’s open-model strategy is not new, but it has accelerated through 2026 as the company faces a different kind of competitive pressure than it did a year ago. Its hardware business depends on developers choosing NVIDIA GPUs and the CUDA software stack, and giving away competitive open-weight models that run best on NVIDIA infrastructure is a way to keep that flywheel spinning even as rivals push their own open-source alternatives. That dynamic has grown more pointed as Chinese labs have expanded their open-weight footprint; this site covered how a wave of Chinese open-source LLMs is reshaping AI infrastructure spending plans earlier this year, and NVIDIA’s Nemotron Coalition reads in part as a US-anchored counterweight to that shift. ZDNet’s own framing of the Super VL launch, “another step forward for US open-weight AI,” reflects that same geopolitical subtext.
Market Impact: Enterprise AI Spending and the $4,500 Question
The free-weights-plus-paid-enterprise-license model is the same playbook NVIDIA has used across its software stack for years, and the reported $4,500-per-GPU-per-year figure for NVIDIA AI Enterprise production licensing gives a concrete anchor for what “open” actually costs at scale. A single node running eight GPUs under that license would run $36,000 a year before any cloud or on-prem hardware costs are factored in, which positions Super VL less as a free alternative to closed APIs and more as a vehicle for selling NVIDIA’s broader infrastructure stack. For a growing enterprise AI buyer base that has spent 2026 getting burned by unpredictable per-token closed-API bills, a fixed per-GPU licensing number, paired with the ability to self-host and fully control data residency, is a meaningfully different pitch than anything Google, Anthropic, or OpenAI currently offers with their closed frontier models.
Historical Context: From Early Nemotron Releases to the Coalition Model
NVIDIA’s Nemotron family did not start as a frontier-scale open model line. Earlier Nemotron releases were positioned mainly as fine-tuned, enterprise-friendly variants of other open base models rather than ground-up frontier systems. The shift toward NVIDIA building and coordinating its own large-scale MoE architecture, developed with the Nemotron Coalition rather than a single internal team, marks a change in ambition. NVIDIA is now positioning itself not just as the hardware layer underneath everyone else’s models, but as a direct publisher of frontier-adjacent open weights competing for developer attention against Meta, DeepSeek, Alibaba’s Qwen line, and Zhipu’s GLM family. That is a markedly more assertive stance than NVIDIA held even twelve months ago, when its software releases were almost entirely infrastructure and tooling rather than headline model weights.
What Comes Next
- Expect NVIDIA to publish a fuller, independently reproducible benchmark comparison for Nemotron 3.5 Super VL within weeks, since the current claims rely heavily on NVIDIA’s own selected comparisons rather than third-party leaderboards.
- Watch how Super VL performs across its documented context range: NVIDIA lists up to 1 million tokens, in line with the 1-million-token windows on Gemini 4 Argon and DeepSeek V4.1 Flash, so real-world long-document accuracy rather than the headline window will shape which document-heavy use cases it can serve.
- More post-training case studies like the Mindbeam partnership are likely to surface over the next quarter, since NVIDIA’s entire pitch for Super VL depends on proving the “post-train it yourself” story with real numbers rather than demos.
- Expect continued downward pressure on open-model pricing and licensing terms as NVIDIA, DeepSeek, Alibaba, and Zhipu all compete for enterprise agent workloads simultaneously.
- The gap between closed frontier leaderboard scores and open-model task-specific claims will likely keep narrowing on specialized benchmarks even as it holds steady or widens on general reasoning, reinforcing a split market: closed models for general assistants, open models for narrow, post-trained enterprise agents.
Frequently Asked Questions
What is NVIDIA Nemotron 3.5 Super VL?
It is an open-weight, 120-billion-parameter mixture-of-experts AI model NVIDIA released on October 15, 2026, with roughly 12 billion active parameters per token, built for multimodal, agentic enterprise tasks.
Is Nemotron 3.5 Super VL free to use?
The model weights are open and available through Hugging Face, NIM, OpenRouter, and ModelScope. Enterprise-grade production deployment through NVIDIA AI Enterprise reportedly carries a $4,500-per-GPU-per-year license, separate from the free weights.
What does the “A12B” in the model’s name mean?
It refers to the roughly 12 billion active parameters the model uses per token, out of 120 to 124 billion total parameters, thanks to its mixture-of-experts architecture.
Does Nemotron 3.5 Super VL actually beat Claude Opus 5.5 or GPT-6.1 Sol?
NVIDIA claims a post-trained version of Super VL beats both on specific specialized task benchmarks, according to VentureBeat’s reporting. That is a narrower, company-reported claim, not an independent general-reasoning leaderboard result; public trackers like BenchLeader still rank Claude Opus 5.5 and Gemini 4 Argon ahead on broad benchmarks as of October 15.
What license does Nemotron 3.5 Super VL use?
Reports describe it as using NVIDIA’s OpenMDW license, which reportedly allows commercial use, fine-tuning, and redistribution of the model weights.
What are the Nemotron 3.5 Bolt models?
They are two smaller companion models to Super VL, sized for roughly 4GB and 24GB memory footprints, aimed at edge and desktop deployment rather than data-center inference. NVIDIA released the 24GB Bolt 74B-A9B on October 15, 2026, alongside Super VL, and the 4GB Bolt E4B on October 23, 2026.
How does Nemotron 3.5 Super VL compare to Liquid AI’s LFM2.5-24B-A2B?
LFM2.5-24B-A2B is a much smaller model, with 24 billion total and 2 billion active parameters, built for on-device reasoning and tool-calling with a confirmed 131,000-token context window. Super VL is roughly five times larger in total parameters and targeted at data-center-scale agent deployments, though NVIDIA documents context lengths up to 1 million tokens for Super VL.
Where can developers download Nemotron 3.5 Super VL?
It is available on Hugging Face under NVIDIA’s organization page, as well as through NVIDIA NIM, OpenRouter, ModelScope, and the build.nvidia.com developer portal.
