Four Nemotron models in one day – but the numbers rest on the company's own claims
October 15, 2026 brought not just a model but a whole structure: NVIDIA's 120B flagship Super VL, two Bolt edge models, Multiverse's compressed Comet 60B and Mindbeam's post-trained CUDA-kernel model. Same day. But behind most of it lie self-published claims without published test protocols.
October 15, 2026 was not an ordinary model launch day. NVIDIA released the expansion of its Nemotron 3.5 family — the multimodal, 120-billion-parameter Super VL plus two Bolt models for local execution — and the same day saw two ecosystem launches around them: Multiverse Computing's compressed Comet 60B and Mindbeam's post-trained model and inference engine for CUDA kernels. Four launches, two partners, and an entire "efficiency stack" within 24 hours.
The picture that emerges is not primarily a race over raw benchmark scores, but over who can deliver the most intelligence per gigabyte, per watt and per contract task — from data center down to 4 GB devices. At the same time, nearly every number in this story rests on companies' own disclosures: VentureBeat notes that NVIDIA has not supplied task definitions, sample sizes or the test protocol behind its benchmark chart, and the ecosystem figures are pure press-release claims.
What actually shipped on October 15
The core of the launch is Nemotron 3.5 Super VL, an open, 120-billion-parameter model that NVIDIA says can work across PDFs, charts, tables, images and webpages while following instructions and using tools. It is, in other words, a multimodal model built for agentic workflows, not just chat.
Around the flagship, ZDNET places two edge models: Nemotron 3.5 Bolt 74B-A9B 24GB, intended for local agent use, and Nemotron 3.5 Bolt E4B 4GB for voice agents, on-device agents and interactive characters. The latter is announced but not due out until October 23 — it is not yet available. NVIDIA markets the Bolt series with the claim of the highest intelligence per gigabyte in the family.
The same day, Multiverse Computing launched Comet 60B, a compressed and optimized version of Super VL, in collaboration with NVIDIA. And Mindbeam announced a collaboration covering two layers of the open AI stack: a post-trained model that generates high-performance CUDA kernels, and an inference engine that runs NVIDIA's open models.
The speed claim: 5x faster
NVIDIA describes Super VL in its press release, as relayed by ZDNET, as "the fastest open agentic model in its class — 5x faster than the next best — helping enterprise agents complete high-volume tasks faster, across workflows that span days or weeks."
The claim rests, according to ZDNET, on benchmark tests from Artificial Analysis, in which Super VL is compared against Ling 3.0 Flash and DeepSeek V4.1 Flash. The methodology behind the comparison is not available in what has been published so far, so "5x faster than the next best" should be read as a marketing claim with a named measurement firm behind it — not as an independently verified result.
Post-training as the defense of the 120B model
The most interesting part of NVIDIA's narrative is not the base model, but what enterprises are meant to build on top of it. According to VentureBeat, NVIDIA says dozens of organizations are already post-training Super VL, and a chart shows improvements of 9 to 44 percent on five specialized tasks — with each specialized version scoring above a named Claude or GPT model on exactly that task.
ZDNET highlights the medical insurance claims example: a 44 percent improvement said to have placed the model well above Claude Opus 5. But ZDNET is clear about the crucial nuance: it means the base model did not beat the proprietary frontier models on its own. NVIDIA's point is that enterprises can tailor the open model and achieve highly specific gains.
NVIDIA lifts up named partners and use cases, as reported by VentureBeat:
- Scale AI, for Skydance's finance agent for vendor credit
- EY, for an M&A due-diligence agent
- AfterQuery, for Hebbia's analysis of credit agreements
- AI21 and Alice, for cybersecurity
- Mercor, for coding agents
VentureBeat also names platform players — Atlassian, CrowdStrike, Perplexity, Salesforce and Synopsys — as integrating or building on the model.
Here the methodological caveat matters: VentureBeat notes that NVIDIA has not supplied the task definitions, sample sizes or test protocol behind the chart's results. Without those, the 9–44 percent figures cannot be replicated or calibrated by independent parties.
The efficiency chain in practice: Comet 60B
Multiverse Computing's Comet 60B shows how the efficiency thesis is put into practice. According to the company's press release, Comet 60B halves the total parameter count to 60 billion, retains all three modalities and the same 12 billion active parameters, and improves on the original in some areas — while otherwise "staying within 5 percent of the original model's accuracy" across text, image and video tasks.
It is worth noting what this reveals about Super VL's architecture: a model with 120 billion total parameters but only 12 billion active is a mixture-of-experts model. It is the same design philosophy that makes compression and sparse activation attractive — you carry a lot of capacity, but pay for only a small share of it per token.
Multiverse describes its multi-step CompactifAI process: profiling with tensor networks, compression, "healing" of the model, and GRPO-style reinforcement learning. Finally the model is quantized and delivered in BF16, FP8 and NVFP4 formats, "so that customers can choose the precision that fits their hardware and latency targets," the press release states. All of these figures are the company's own, without independent verification.
Mindbeam: a post-trained model that writes CUDA kernels
Mindbeam's contribution is perhaps the most technically specific. The company has post-trained Super VL into "Litespark Nemotron 3.5 Super VL," a model said to generate high-performance CUDA kernels — the GPU code that makes computations run fast. According to the press release, the model improved on KernelBench Level 2 from 22.8 to 82.1 percent, which the company says surpasses GPT 5.6 Sol's 57 percent.
Mindbeam also cites first-class inference support in its own engine, at 962 tokens per second per user on two H200 GPUs. The same caveat applies here: these are self-reported figures from a press release with no published protocol or independent replication. But they point to a real technical trend — models trained specifically to produce system code, and inference engines optimized for particular model families.
The family context: from Lightning to 4 GB
The October 15 expansion sits in a longer line. Nemotron 3.5 Lightning, a mixture-of-experts model with 30 billion total parameters but only 3 billion active at a time, was launched and documented on August 11, 2026, aimed at long-running agents with low inference costs.
With the Bolt models, the "highest intelligence per gigabyte" formula is pushed down to 24 GB and further to 4 GB — models that can run locally on desktops and in devices, for voice agents and interactive characters. The family's spread is thus deliberate: a 120B flagship for enterprise post-training, 30B Lightning for long-running agents, and Bolt models for the edge.
The verification gap
There is no primary NVIDIA documentation — model cards, technical report or the announcement itself — in the public record this coverage is built on. All specifications and performance claims about Super VL and Bolt come via VentureBeat, ZDNET or partners' press releases. Concretely, that means:
- The benchmark chart (9–44 percent) lacks task definitions, sample sizes and a test protocol, according to VentureBeat.
- The 5x speed claim rests on Artificial Analysis tests as NVIDIA itself cites them.
- Comet 60B's "within 5 percent" accuracy and Mindbeam's KernelBench and throughput figures are pure self-published claims.
None of this necessarily means the numbers are wrong — but they are all structurally claims from parties with an interest in the outcome, and they should be verified before being treated as fact.
Open questions
Three things could change the picture. First: independent replication of the benchmark chart and of the Multiverse and Mindbeam figures. Second: whether Bolt E4B 4GB actually lands on October 23 as promised, and how it performs in practice on 4 GB hardware.
Third: Shattered.io references a rumor, relayed via The Information, that NVIDIA is working on a Nemotron 4 successor with roughly a trillion parameters. This is unconfirmed — neither launch date nor parameter count has been verified — and should be treated as exactly what it is: a rumor.
What the day tells us
The most significant thing about October 15 may not be the models themselves, but the speed. An open flagship launched in the morning was, by the end of the day, compressed to half its parameter count by one company and post-trained to write GPU code by another. Open-weight ecosystems now take shape in hours, not months — which makes both the launch and the caveats relevant at the same time. The faster these ecosystems grow, the more important it becomes to keep an eye on the distance between press-release numbers and verified results.