Nvidia used October 15 to widen its open-weight model lineup again, confirming three new additions to its Nemotron family that stretch from a 120-billion-parameter vision-language flagship down to a 4GB model built for pocket-sized hardware. The announcement, detailed in an Nvidia investor release and picked up by outlets including Yahoo Finance, MarkTechPost, and Unite.AI, lands at a moment when every major AI lab is racing to put open weights into as many hands as possible before rivals do.
The three models are Nemotron 3.5 Super VL, Nemotron 3.5 Bolt 74B-A9B 24GB, and Nemotron Bolt E4B 4GB. Together they cover a deployment range Nvidia says spans data-center systems all the way to desktop computers, a span the company has been building toward since it first opened up the Nemotron line. This piece looks at what each model actually targets, how the release fits into Nvidia’s wider open-source push, and why the timing matters for a chip market that just spent a week absorbing a broader semiconductor selloff.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Nvidia Announced on October 15
According to the Nvidia investor relations announcement, the company is expanding its open model families to support what it frames as the next wave of agentic, physical, and healthcare AI. The release names three specific models rather than a single flagship, a pattern Nvidia has leaned on before with Nemotron but rarely at this spread of sizes in one sitting.
Reports describe the lineup as designed to run from data-center systems to desktop computers, which is a wider range than most single model families attempt to cover. Rather than one model scaled down through quantization, Nvidia shipped three distinct architectures, each tuned for a different slice of that range. No pricing was disclosed in the available reporting, and Nvidia’s materials do not include a named executive quote describing the strategy, so this analysis sticks to the confirmed specs and deployment framing rather than speculating on intent beyond what the company stated.
Nemotron 3.5 Super VL: A 120-Billion-Parameter Vision-Language Model
Nemotron 3.5 Super VL is the largest of the three, built at 120 billion parameters as an open vision-language model. That places it in the same general weight class as other frontier-scale open releases this year, though Nvidia has not published benchmark scores alongside the announcement, so direct performance claims remain unconfirmed pending independent testing.
Vision-language models handle both images and text in a single reasoning pass, which is the capability Nvidia is leaning on for healthcare and physical AI use cases, two categories the announcement explicitly calls out. A 120-billion-parameter open model aimed at multimodal reasoning is a heavier lift than a text-only model of the same size, since it needs to encode visual representations alongside language ones. Nvidia’s own earlier Nemotron 3 Ultra, which launched at 550 billion parameters, shows the company is comfortable shipping large open weights, but Super VL’s 120B footprint suggests a deliberate trade-off toward something closer to practical, deployable scale rather than a pure parameter-count flex.
Nemotron 3.5 Bolt 74B-A9B 24GB: Built for Agentic Tasks at the Edge
The middle model in the trio, Nemotron 3.5 Bolt 74B-A9B 24GB, is described in the reporting as an edge model built for general agentic tasks, including navigating web browsers and operating applications. The 74B-A9B naming follows a convention common to mixture-of-experts open models, where the larger figure typically represents total parameters and the smaller one an active subset used per inference pass, though Nvidia’s public materials do not spell out the architecture in that level of detail.
What is confirmed is the 24GB memory footprint, which puts this model within reach of a single high-end consumer GPU or a modest workstation setup rather than requiring a server rack. That matters for the agentic use case Nvidia is targeting. A model that needs to click through a browser, fill out a form, or operate a desktop application benefits from running close to the device it is controlling, cutting the latency that comes from round-tripping every action to a distant data center.
Nemotron Bolt E4B 4GB: The Smallest Model in the Family
At the opposite end sits Nemotron Bolt E4B 4GB, pitched as an edge model for physical and conversational agents. A 4GB footprint is small enough to run on hardware well below a typical gaming laptop, which opens the door to robotics controllers, smart home devices, and other embedded systems that cannot spare the power budget or memory a larger model demands.
Pairing “physical” and “conversational” in the same product description signals Nvidia is positioning this model for robots and voice-driven devices alike. A warehouse robot that needs to interpret a spoken instruction and translate it into a movement plan, or a smart speaker handling a back-and-forth conversation, both fall under that umbrella. The company has not published specific robotics partners or shipping hardware tied to E4B in the available reporting, so any claim about where it will actually run first stays speculative for now.
Comparing the Three New Nemotron Models
Laid out side by side, the size gap between the three new releases is the most striking feature of the announcement. Nvidia is not offering one model at three quantization levels. It shipped three separate products aimed at three different hardware tiers.
| Model | Size | Primary Target Use | Deployment Tier |
|---|---|---|---|
| Nemotron 3.5 Super VL | 120 billion parameters | Vision-language reasoning for healthcare and physical AI | Data center / cloud |
| Nemotron 3.5 Bolt 74B-A9B | 24GB memory footprint | General agentic tasks: browsers, apps | Edge / high-end desktop |
| Nemotron Bolt E4B | 4GB memory footprint | Physical and conversational agents | Edge / embedded / robotics |
That spread from a cloud-scale vision model to a 4GB edge model is the real story of October 15, more than any single spec. It tells developers that Nvidia wants the Nemotron brand attached to every point on the deployment curve, not just the frontier end where most open-weight headlines get made.
Why Nvidia Keeps Betting on Open Weights
Nvidia’s open-model push did not start this week. The company has spent much of 2026 building out a Nemotron lineup that now spans multiple generations, and it backed that effort with a reported acquisition of Hugging Face, the open-model hosting platform that counts millions of developers among its users. Owning distribution infrastructure alongside the models themselves gives Nvidia a direct line to the developers who decide which open weights actually get adopted.
The strategic logic is straightforward even without a named executive quote to confirm it. Every model that runs well on Nvidia hardware, whether that hardware sits in a data center or a desktop, reinforces demand for Nvidia GPUs. An open model is also a hedge against closed competitors locking developers into their own ecosystems. If a company can run Nemotron 3.5 Super VL on its own infrastructure instead of paying per-token fees to a closed API, Nvidia still captures the hardware spend even if it gives away the weights for free.
From Data Center to Desktop: The Deployment Range Strategy
The explicit framing in Nvidia’s announcement, that these models run from data-center systems to desktop computers, lines up with a hardware push the company has been making in parallel. Nvidia’s RTX Spark laptop line, which launched with up to 128GB of memory aimed squarely at local AI workloads, gives buyers a machine that could plausibly run the Bolt 74B-A9B model without touching a cloud API.
That is not a coincidence. Nvidia sells the chips that run these models at every tier it is now releasing open weights for. A 120B vision-language model needs data-center-grade GPUs. A 24GB edge model fits a high-end RTX card. A 4GB model can run on far more modest silicon, including some of Nvidia’s smaller embedded and automotive-grade chips. Shipping open weights tuned for each tier gives developers a reason to buy Nvidia hardware at whatever scale they are building for, rather than defaulting to a single cloud provider’s closed model and whatever chip that provider happens to use on the back end.
How This Fits Nvidia’s Broader Nemotron Timeline
Nemotron has moved through several distinct phases this year. Nvidia’s earlier Nemotron 3 Ultra arrived at 550 billion parameters as a dense, large-scale open release, but reports at the time noted it trailed some closed rivals on benchmark comparisons. The company followed that with Nemotron 3.5 Lightning, a smaller, free 30B model aimed at developers who wanted something runnable without massive infrastructure, alongside a push to get Nemotron onto AWS Bedrock as part of a multi-million-GPU cloud deal.
October 15’s three-model drop reads as the next step in that same arc: instead of one model trying to serve every use case, Nvidia is now shipping purpose-built variants across the entire size spectrum simultaneously. That is a different approach than iterating on a single flagship release after release, and it suggests Nvidia has concluded that the open-weight market rewards breadth of deployment options as much as it rewards raw benchmark leadership.
The Competitive Landscape: Mistral, Liquid AI, DeepSeek, and the Open-Weight Race
Nvidia is not the only company racing to put large open models into developers’ hands this quarter. Mistral’s Large 4 launched at a reported 1.05 trillion parameters, priced at $1.36 per million tokens, aiming squarely at enterprise customers who want a flagship-scale model without a closed API’s recurring costs. Liquid AI took the opposite approach with LFM2.5, a 24-billion-parameter model built for efficiency, claiming 184 tokens per second in throughput testing. Chinese labs have kept pace too: DeepSeek’s V4.1 Flash API pushed a 1-million-token context window, while comparisons between GLM 5.3 Prime, Qwen3.8 Max, and Step 5 have shown benchmark gaps as wide as 39 points between leading open-weight contenders.
| Open-Weight Family | Reported Scale | Positioning |
|---|---|---|
| Nvidia Nemotron 3.5 | 4GB to 120B parameters across three models | Edge-to-data-center deployment range |
| Mistral Large 4 | 1.05 trillion parameters | Enterprise flagship, $1.36 per million tokens |
| Liquid AI LFM2.5 | 24 billion parameters | Efficiency-focused edge model, 184 tokens/sec |
| DeepSeek V4.1 Flash | 1 million token context window | API-first, long-context tasks |
| GLM 5.3 Prime / Qwen3.8 Max / Step 5 | Up to 39-point benchmark spread reported between them | Competing Chinese open-weight tier |
What separates Nvidia’s play from most of this field is that it is not trying to win on a single benchmark leaderboard. Mistral and the Chinese labs are largely competing on flagship scale and raw capability. Nvidia’s three-model drop instead competes on range, covering a spread no single rival model family currently spans in one announcement. That is a harder thing to measure with a leaderboard, but it may matter more to developers choosing a model family to standardize on for products that need to run everywhere from a server to a handheld device.
Market Reaction: Nvidia Stock Holds Steady Amid a Chip Sector Selloff
The Nemotron announcement landed during a rough stretch for chip stocks broadly. A Yahoo Finance report noted Nvidia bucking a premarket selloff that hit Micron, Intel, and AMD shares, even as the wider sector absorbed a rout tied to memory pricing and PC chip demand concerns. Nvidia’s relative resilience that week lines up with a pattern seen elsewhere this quarter, where Nvidia has edged higher while Intel and AMD slide on PC chip-related news.
It would be a stretch to credit three open model releases with moving a stock with Nvidia’s market capitalization on their own. Data-center GPU demand, AI capital expenditure guidance, and memory supply dynamics all carry far more weight on any given trading day. But the timing is a useful reminder that Nvidia’s software and model releases are not disconnected from its hardware story. Every open model it ships is also a small advertisement for the chips that run it best.
What Developers and Enterprises Should Watch For
Nvidia’s announcement did not include pricing, exact licensing terms, or a public release date for general availability of all three models, so developers evaluating them should treat specifics beyond the confirmed parameter counts and memory footprints as pending until Nvidia or Hugging Face publish model cards. A few practical questions are worth tracking as the rollout continues.
- Which license each model ships under, since that determines whether commercial deployment is unrestricted or requires a separate agreement
- Whether independent benchmark results confirm Nvidia’s implied performance claims for Super VL’s multimodal reasoning
- How quickly the models land on Hugging Face and other model hubs for direct download
- Whether Nemotron Bolt E4B ships with reference hardware or robotics partner integrations
- How the Bolt 74B-A9B model’s real-world memory use compares to the stated 24GB figure once quantized for consumer GPUs
Teams already experimenting with Nvidia’s earlier open releases have a reference point for what to expect. The company’s guide to running Nemotron 3.5 Lightning locally walked through a 12-step local setup for that smaller model, and a similar pattern is likely once Bolt E4B and Bolt 74B-A9B reach wider availability. For now, the safest move for most teams is to wait for official model cards rather than build against specs pulled from secondary reporting.
A Generic Pattern for Loading Open-Weight Models
Nvidia has not published code samples or a Hugging Face repository identifier for any of the three new models in the available reporting, so the snippet below is illustrative only. It shows the general pattern developers typically follow to load an open-weight model once it appears on a model hub, not a confirmed command for Nemotron 3.5 specifically.
from transformers import AutoModelForCausalLM, AutoTokenizer
# Replace with the official model identifier once Nvidia publishes it
model_id = "nvidia/PLACEHOLDER-nemotron-3.5-bolt-e4b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
inputs = tokenizer("Describe the task the agent should perform:", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))Once Nvidia or Hugging Face publish the real repository paths, that placeholder line is the only part of this pattern that changes for most standard text and agentic workloads.
Historical Context: How Open Weights Became Nvidia’s Second Front
Two years ago, Nvidia’s public identity was almost entirely about chips. The company sold the hardware that trained and ran AI models, and it largely left the model-building to customers like OpenAI, Anthropic, Meta, and Google. That changed as Nvidia built out a full Nemotron family, reportedly acquired Hugging Face, and started treating open weights as a core part of its product strategy rather than a side project.
The reasoning behind that shift is not hard to follow. Every open model is a magnet for developer attention, and developer attention is the thing that decides which chips get bought at scale. A closed competitor’s model running on a closed competitor’s cloud infrastructure does nothing for Nvidia’s hardware sales. An open model that a developer downloads and runs on their own GPU cluster, desktop workstation, or robotics controller keeps Nvidia in the loop no matter who built the application layer on top.
Predictions: What Comes Next for Nvidia’s Open Model Strategy
Based on the pattern Nvidia has followed across its Nemotron releases this year, a few directions look likely over the coming months, though none of these should be read as confirmed Nvidia plans.
- Nvidia will likely publish independent benchmark results for Nemotron 3.5 Super VL within weeks of release, since every prior Nemotron launch has followed with comparison data against rival open models
- Expect the Bolt 74B-A9B and Bolt E4B models to show up bundled into Nvidia’s own developer tooling and RTX Spark software stack before third-party integrations catch up
- Robotics and physical-AI partners are the most plausible early adopters of Nemotron Bolt E4B, given Nvidia’s existing push into that market through its broader AI hardware lineup
- Competing labs, particularly Mistral and the Chinese open-weight players, are likely to respond with their own edge-tier releases rather than only competing at flagship scale
- Pricing for any hosted or enterprise-support version of these models will probably arrive separately from the open-weight release itself, following the pattern Nvidia set with earlier Nemotron tiers
Why the Edge-to-Cloud Range Matters More Than the Headline Parameter Count
It is tempting to read this announcement as another entry in the ongoing parameter-count arms race, where Super VL’s 120 billion parameters becomes the headline and the two Bolt models become footnotes. That framing misses what actually makes this release distinct. A single company shipping a coherent model family that runs from a data center down to a 4GB footprint is rarer than another large flagship model landing on a leaderboard.
Most developers do not need a 120-billion-parameter model for most tasks. They need something that runs reliably on the hardware they already have, whether that is a cloud GPU instance, a gaming laptop repurposed for local inference, or an embedded board controlling a physical device. Nvidia shipping all three tiers on the same day, under the same family name, gives developers a reason to standardize on Nemotron across an entire product line rather than mixing and matching model families as hardware requirements change project to project.
Frequently Asked Questions
What did Nvidia announce on October 15, 2026?
Nvidia confirmed three new open models in its Nemotron family: Nemotron 3.5 Super VL, a 120-billion-parameter vision-language model, Nemotron 3.5 Bolt 74B-A9B with a 24GB memory footprint for edge agentic tasks, and Nemotron Bolt E4B, a 4GB model for physical and conversational agents.
Are these models free to use?
Nvidia’s available reporting does not state pricing or licensing terms for the three new models. Developers should wait for official model cards from Nvidia or Hugging Face before assuming free commercial use.
What hardware do I need to run Nemotron 3.5 Super VL?
As a 120-billion-parameter vision-language model, Super VL is built for data-center or cloud-scale GPU infrastructure, not consumer hardware, according to Nvidia’s deployment range description.
Can Nemotron Bolt 74B-A9B run on a gaming laptop?
Its stated 24GB memory footprint fits within the range of Nvidia’s own RTX Spark laptops and similar high-end consumer or workstation GPUs, making it the most locally deployable of the three new models aside from the smaller E4B variant.
What is Nemotron Bolt E4B used for?
Nvidia describes it as an edge model for physical and conversational agents, meaning robotics controllers and voice-driven devices are the most likely early use cases, though no specific shipping hardware has been confirmed yet.
How does this compare to Nvidia’s earlier Nemotron 3 Ultra?
Nemotron 3 Ultra launched earlier at 550 billion parameters as a single large dense model. The October 15 release takes a different approach, shipping three distinct models across a wider size range instead of one flagship.
Did this announcement affect Nvidia’s stock price?
Nvidia shares held up relative to peers like Micron, Intel, and AMD during a broader chip sector selloff the same week, according to Yahoo Finance, though that move was tied more closely to memory and PC chip demand concerns than to the Nemotron release itself.
How does Nemotron 3.5 compare to Mistral Large 4 or Liquid AI’s LFM2.5?
Mistral Large 4 targets enterprise flagship use at 1.05 trillion parameters, while Liquid AI’s LFM2.5 focuses on efficiency at 24 billion parameters. Nvidia’s three-model Nemotron 3.5 release instead spans both ends of that spectrum and the data-center tier above it in a single coordinated announcement.
