Nvidia spent the past three years convincing the world it makes the chips that power AI. On June 4, 2026, it made a different kind of statement: it gave away the model too. Nvidia Nemotron 3 Ultra, a 550-billion-parameter open-weight language model, is now sitting on Hugging Face for anyone to download, fine-tune, or deploy, no API key required.
That alone would be news. A company with a market capitalization built almost entirely on selling GPUs just shipped one of the largest open-weight models on the planet, free to use under a permissive license. But the more interesting part of the story happened seven weeks later, when Nvidia CEO Jensen Huang used his first-ever post on X to throw his company’s weight behind an industry letter defending open-weight AI models from potential US restrictions. Read together, the two events describe a company that no longer sees itself as just a hardware vendor. Nvidia is now an active participant in deciding which AI models get built in the open, and it is spending real engineering and political capital to keep it that way.
This piece breaks down what Nemotron 3 Ultra actually is, how it stacks up against Claude Opus 4.8, GPT-5.5 and other 2026 frontier models, why Nvidia is willing to give away a model that likely cost tens of millions of dollars to train, and what the broader open-weight policy fight signals for the rest of 2026.
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 Is Nemotron 3 Ultra? Nvidia’s Largest Open Model Yet
Nemotron 3 Ultra is a 550-billion-parameter model built on a hybrid Mamba-Attention Mixture-of-Experts (MoE) architecture, with only about 55 billion parameters active per token, according to NVIDIA Research’s June 2026 technical report and Nvidia’s developer blog. Nvidia’s own NIM model card confirms the same specs, listing a formal release date of June 4, 2026 and NVFP4-quantized weights for the 550B-A55B configuration, alongside the standard BF16 build covered in the technical report. That means roughly one-tenth of the network switches on for any given token, a sparsity ratio Nvidia is using to justify a 1-million-token context window without the runaway inference costs a dense model that size would carry.
The model launched on June 4, 2026 as the flagship of the Nemotron 3 family, which also includes smaller Nano and Super variants. Nvidia frames Ultra as an “open frontier-reasoning and orchestration model” aimed at agentic workloads: long-running tasks, multi-step tool use, and coding agents that need to hold large amounts of context without losing track of earlier steps. Weights are published on Hugging Face in both BF16 and NVFP4 (Nvidia’s 4-bit floating point format) under a permissive open license Nvidia calls OpenMDW-1.1, which secondary coverage describes as following Linux Foundation-style terms rather than a research-only license.
Unlike a closed model such as Claude Opus 4.8 or GPT-5.5, there is no waitlist and no per-token invoice from Nvidia. Anyone with enough GPU memory, or a rented cluster, can run it themselves.
Inside the Architecture: What Hybrid Mamba-Attention MoE Actually Means
Most 2024-era large language models were pure transformers: every token attends to every other token, which gets expensive fast as context length grows. Nemotron 3 Ultra instead interleaves three layer types, according to Nvidia’s technical report. Mamba-2 state-space layers process sequences with a fixed-size memory state instead of full attention. Traditional attention layers stay in the mix for tasks that need precise long-range lookups. MoE layers, using a design Nvidia calls LatentMoE, route each token to a small subset of specialized expert sub-networks.
Nvidia also built in multi-token prediction (MTP), a technique that lets the model draft several tokens ahead in a single forward pass and verify them, cutting the number of sequential steps needed to generate a response. The combined effect shows up in one benchmark in particular: RULER, a long-context retrieval test, where Nemotron 3 Ultra scored 94.7% at 1 million tokens, per Nvidia’s published long-context retrieval result in the Nemotron 3 Ultra materials. Independent testing backed up the speed claims: a June 2026 Computex recap from ExplainX.ai clocked the model at more than 300 tokens per second in real-world throughput, a figure Nvidia attributes to roughly 90% of the model’s parameters sitting dormant on any given token. That combination, high throughput despite the model’s size, is the number Nvidia is most eager to have compared against competitors, since it is the one place the hybrid architecture’s design goal, cheap long-context reasoning, shows up most directly.
Benchmark Scores: Where Nemotron 3 Ultra Actually Lands
Nvidia published a full benchmark table alongside the release, and it’s worth reading closely because it does not show Nemotron 3 Ultra beating the leading closed models. On the Artificial Analysis Intelligence Index, a composite score spanning reasoning, coding, knowledge and agentic tasks, Nemotron 3 Ultra scored 48.2 in BF16 and 47.7 in the more efficient NVFP4 format. Independent benchmarking outfit Vals.ai reached a similar conclusion from a different angle: as of August 2026, its Vals Index ranked Nemotron 3 Ultra 5th among open-weight models with a 43.99% score. On coding-specific tests, it performs more competitively: 71.9% on SWE-bench Verified and 67.8% on SWE-bench Multilingual in BF16.
| Benchmark | BF16 Score | NVFP4 Score | What It Measures |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 48.2 | 47.7 | Composite reasoning, coding, knowledge, agentic tasks |
| SWE-bench Verified | 71.9% | 69.7% | Real-world GitHub issue resolution |
| SWE-bench Multilingual | 67.7% | 65.8% | Coding across multiple programming languages |
| Terminal-Bench 2.1 | 56.4% | 53.9% | Agentic command-line task completion |
| GPQA Diamond | 87.0% | 87.9% | Graduate-level science reasoning |
| Humanity’s Last Exam | 26.7% | 26.1% | Frontier-difficulty general knowledge |
| IFBench | 81.7% | 82.3% | Instruction-following accuracy |
| RULER (1M context) | 94.7% | N/A | Long-context retrieval accuracy |
The pattern is consistent: strong on agentic and long-context tasks, solid on coding, but well behind frontier closed models on the hardest general-reasoning tests like Humanity’s Last Exam, where a 26.7% score puts real distance between Nemotron 3 Ultra and the models it is implicitly being compared to. That mid-pack positioning shows up on other third-party leaderboards too: BenchLM.ai’s BenchAlign ranking placed Nemotron 3 Ultra 154th out of 221 models tracked, with a composite score of 46.53 out of 100, as of August 2026.
Nemotron 3 Ultra vs Claude Opus 4.8 and the Frontier Field
Context matters here, because the Artificial Analysis Intelligence Index is the same yardstick Claude Opus 4.8 used to take the top spot in June 2026 at a score of 61.4. GPT-5.5 in its highest-effort “xhigh” configuration scored 60.2 on the same index, and OpenAI’s newer GPT-5.6 Sol variant landed at 59 in its maximum-effort mode. Nemotron 3 Ultra’s 47.7-48.2 score puts it roughly 13 points behind the closed frontier, and it also trails at least one open-weight rival: Moonshot AI’s Kimi K2.6 scored 53.9 on the same index, about six points ahead of Nvidia’s new flagship. A separate ranking from DigitalApplied, tracking the same launch-week benchmark data, placed Nemotron 3 Ultra’s Intelligence Index score of 48 9th out of 89 models evaluated as of June 2026, a reminder of just how crowded the open-weight field it’s competing against has gotten.
| Model | Artificial Analysis Intelligence Index | Access Model |
|---|---|---|
| Claude Opus 4.8 | 61.4 | Closed (API) |
| GPT-5.5 (xhigh) | 60.2 | Closed (API) |
| GPT-5.6 Sol (max) | 59.0 | Closed (API) |
| Kimi K2.6 | 53.9 | Open-weight |
| Nvidia Nemotron 3 Ultra (BF16) | 48.2 | Open-weight |
| Nvidia Nemotron 3 Ultra (NVFP4) | 47.7 | Open-weight |
That gap is worth sitting with. Nvidia did not ship the smartest model on the planet, and it likely never intended to. What it shipped is the largest fully open model with a legitimate claim to frontier-adjacent performance on agentic and coding tasks, released by a company whose actual business is renting out the hardware needed to run it. Whether Nemotron 3 Ultra “wins” the leaderboard matters less to Nvidia’s bottom line than whether developers choose to build on it, and building on it means buying or renting more Nvidia compute.
From 15B to 550B: Nvidia’s Nemotron Family Growth
Nemotron 3 Ultra is not Nvidia’s first open model, but it represents a scale jump from anything the company has shipped before. Nvidia’s open Nemotron lineage stretches back to Nemotron-4 15B and the 340-billion-parameter Nemotron-4 in 2024, both dense models aimed mostly at synthetic data generation and enterprise fine-tuning rather than frontier reasoning.
| Model | Total Parameters | Active Parameters | Released |
|---|---|---|---|
| Nemotron-4 15B | 15B | Dense (all active) | 2024 |
| Nemotron-4 340B | 340B | Dense (all active) | 2024 |
| Nemotron 3 Nano | ~30B | Up to 3B | 2026 |
| Nemotron 3 Super | ~100-120B | Up to 10-12B | 2026 |
| Nemotron 3 Ultra | 550B | Up to 55B | June 4, 2026 |
The shift from dense to sparse MoE architecture between the 2024 and 2026 generations is the more telling change. Nvidia’s 2024 models activated every parameter on every token, the brute-force approach. The 2026 Nemotron 3 family activates roughly 10% of its total parameters per token across all three sizes, a bet that sparsity, not raw scale, is what makes open-weight models economically viable to actually run at the sizes labs now want to build them.
Why Nvidia Keeps Giving Away Its Biggest Models
The obvious question is why a hardware company keeps spending compute to train models it then hands out for free. Nvidia has never published a line-item explanation, but the pattern lines up with a strategy the company has run since the first Nemotron release: open weights pull developers, researchers and enterprises deeper into Nvidia’s software stack, including CUDA, the NeMo framework, and NIM (Nvidia Inference Microservices), all of which run best on Nvidia’s own hardware. A developer who fine-tunes Nemotron 3 Ultra for a specific task is, by definition, a customer for the GPUs or cloud instances needed to serve it.
Jensen Huang made the underlying philosophy explicit weeks after the Nemotron 3 Ultra launch, when he used his first public post on X to back an industry-wide defense of open-weight models:
“The world needs both closed frontier models and open frontier models.”
Jensen Huang, CEO, NVIDIA, quoted by The Register
That framing, both/and rather than either/or, is consistent with a company that sells the picks and shovels regardless of which camp wins. Nvidia does not need Nemotron 3 Ultra to outscore Claude Opus 4.8. It needs the open-weight ecosystem to keep growing, because every open model that gets fine-tuned, deployed and scaled by a startup or enterprise is inference workload that has to run somewhere, and Nvidia wants that somewhere to run on its chips.
The Open Weights and American AI Leadership Letter
The Nemotron 3 Ultra launch and Huang’s July social media post are two parts of the same story. On July 24, 2026, a coalition of technology companies published a letter titled “Open Weights and American AI Leadership,” organized with Microsoft and posted on Microsoft’s corporate site, according to Politico’s coverage of the launch. Huang amplified it publicly the same day. The letter initially carried roughly 25 signatories, including Nvidia, Microsoft, Meta and Palantir, and grew to about 50 organizations by July 26, with later reporting noting that OpenAI and Google were added to the list after the initial launch, per Forbes.
The letter’s concrete asks are policy-oriented rather than symbolic: expand compute access for startups and academic researchers, invest in shared training infrastructure such as public datasets and evaluation frameworks, and avoid what the signatories call “premature restrictions” on open-weight models. Politico’s reporting frames the timing as a response to a live debate inside the White House over whether to restrict or ban Chinese open-weight models on national security grounds, a policy move that, if broadened, could sweep in restrictions on open-weight releases generally, including Nvidia’s own.
Forbes’ reporting also flags who did not sign: Amazon and Anthropic were both absent from the expanded list, a detail that underscores the letter is not a full-industry consensus so much as a coalition of companies with a specific commercial interest in open weights remaining unrestricted.
What Industry Voices Are Saying
Huang’s public comments around the letter go further than a routine corporate statement, arguing that open models are a matter of both competitiveness and national strategy:
“AI will transform every industry, power every company, and be built by every country.”
Jensen Huang, CEO, NVIDIA, quoted by CNBC
In his original post sharing the letter, Huang was more direct about the stakes he sees for safety and access, not just competition:
“For my first post, I’m sharing a letter @NVIDIA signed on why open models matter. AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.”
Jensen Huang, CEO, NVIDIA, posted on X
The letter’s text itself, signed collectively rather than attributed to one executive, lays out the core economic argument the signatories are making to policymakers:
“Open-weight models are an important part of that foundation because they make advanced AI more accessible, adaptable, and widely available.”
Open Weights and American AI Leadership letter, quoted by Business Insider
Taken together, the quotes describe a coordinated lobbying push, not an isolated Nvidia PR moment. Nemotron 3 Ultra gives Nvidia a concrete product to point to when it argues that open models are safe, useful and worth protecting from regulation.
Where and How to Access Nemotron 3 Ultra
Developers have three main entry points. The model weights, in both BF16 and NVFP4, are published on Hugging Face for direct download and self-hosting. Nvidia also lists the model on its own build.nvidia.com catalog with a hosted model card for testing through NIM, Nvidia’s inference microservice packaging. And third-party routers, including OpenRouter, list the model for pay-per-token access without requiring users to provision their own GPU infrastructure.
Running the BF16 checkpoint at full precision requires substantial GPU memory given the 550-billion parameter count, even with only 55 billion active per token, since all weights still need to be resident in memory for MoE routing. The NVFP4 quantized version exists specifically to lower that bar, trading a small amount of benchmark accuracy (generally under 2 points across the table above) for a meaningfully smaller memory footprint.
Quick Start: Calling Nemotron 3 Ultra Through an OpenAI-Compatible API
For developers who don’t want to self-host, routers exposing Nemotron 3 Ultra through an OpenAI-compatible interface let you swap it into existing code with a model name change. A generic example against that kind of endpoint looks like this:
import requests
response = requests.post(
"https://openrouter.ai/api/v1/chat/completions",
headers={
"Authorization": "Bearer YOUR_API_KEY",
"Content-Type": "application/json",
},
json={
"model": "nvidia/nemotron-3-ultra-550b-a55b",
"messages": [
{"role": "user", "content": "Summarize the last three commits in this repo and flag any breaking changes."}
],
"max_tokens": 1024,
},
)
print(response.json()["choices"][0]["message"]["content"])The same pattern applies to any OpenAI-compatible client library, since most routers implement that interface rather than requiring a bespoke SDK. For self-hosted deployment, Nvidia’s GitHub repository includes a usage cookbook covering setup with its NeMo framework.
A Crowded Month: MiniMax M3, Kimi K2.7 and GLM-5.2 Also Shipped
Nemotron 3 Ultra did not launch into an empty field. It landed in the middle of one of the busiest two-week stretches for open-weight releases in 2026. MiniMax shipped MiniMax M3 on June 1, an open-weight model with a 1-million-token context window and a reported 59.0% score on SWE-bench Pro. Moonshot AI followed on June 12 with Kimi K2.7 Code, a 1-trillion-parameter MoE model built specifically for coding workloads and reportedly using 30% fewer “thinking tokens” than its predecessor to reach similar answers. Zhipu AI’s GLM-5.2 arrived June 13 with its own 1-million-token context window and tiered “High” and “Max” effort reasoning modes.
Four major open-weight releases inside two weeks, from four different companies on two different continents, is not a coincidence so much as a sign of where competitive pressure in AI has shifted. A year ago, the open-weight conversation centered on whether any open model could approach frontier closed-model quality at all. By June 2026, the question has moved to differentiation: long context, coding specialization, efficiency, and now, with Nvidia’s entry, direct hardware-vendor backing.
Market Impact: The Open-Weight Race Is Reshaping AI Compute Demand
The business logic behind Nvidia publishing Nemotron 3 Ultra becomes clearer once you separate “who has the smartest model” from “who captures the inference spend.” Every open-weight model released this year, regardless of which lab built it, still needs somewhere to run. A 550-billion-parameter MoE model with a 1-million-token context window is not something most organizations will run on a laptop. It demands multi-GPU clusters, exactly the hardware Nvidia sells.
That dynamic gives Nvidia a reason to support open weights even when its own model doesn’t top the leaderboard. If a competitor’s open model wins on benchmarks but developers deploy it on Nvidia GPUs anyway, because that’s where the software ecosystem and tooling are most mature, Nvidia still wins the underlying compute contract. Publishing its own frontier-adjacent open model adds a second path: developers who standardize directly on Nemotron 3 Ultra get first-party support, NIM packaging, and a company motivated to keep improving the model precisely because doing so sells more hardware.
The policy fight adds a second layer of market impact. If the White House had moved to broadly restrict open-weight model distribution, whether aimed at Chinese labs specifically or open-weight releases more generally, it would have threatened the exact ecosystem Nvidia is counting on to drive demand for its chips outside the small circle of companies that can afford frontier closed-model training runs. The letter is best read as Nvidia and its co-signatories protecting a demand channel, not just defending an abstract principle.
Historical Context: Open Weight vs. Closed Frontier Models
The open-versus-closed debate in AI is not new, but its center of gravity has moved twice in three years. Meta’s Llama 2 release in mid-2023 is generally credited with proving that a credible open-weight alternative to closed frontier models was commercially viable, though it trailed GPT-4-class models by a wide margin at the time. The second inflection came when a Chinese lab’s reasoning-focused open release in early 2025 briefly rattled US AI stocks by showing that a much smaller training budget could produce a model competitive with far more expensive closed systems, intensifying every major lab’s interest in publishing open variants alongside flagship closed models.
By mid-2026, nearly every major AI lab, American and Chinese, now ships some open-weight tier: Meta’s Llama family, Alibaba’s Qwen series, DeepSeek’s releases, Zhipu’s GLM line, Moonshot’s Kimi models, MiniMax, and now Nvidia’s Nemotron 3 family. What’s genuinely new about Nvidia’s entry is that a chip and systems company, rather than a research lab, is now shipping frontier-adjacent open weights as a strategic complement to its hardware business. That’s a different motive structure than a research lab publishing open weights for reputation or safety research, and it suggests open-weight releases are becoming as much an infrastructure-demand play as an AI-capability one.
5 Predictions for Nvidia’s Open Model Strategy Through Late 2026
Based on Nvidia’s release cadence, the current benchmark gap, and the political dynamics around the Open Weights letter, here’s how this is likely to play out over the rest of the year:
- The Open Weights letter’s signatory list keeps growing. With roughly 50 organizations on board by late July and the underlying policy debate over restricting open models still unresolved, expect additional signatories through the rest of 2026, particularly if the White House moves closer to a formal decision on Chinese open-weight models.
- Nvidia will iterate on Nemotron 3 rather than immediately jump to a Nemotron 4 flagship. The two-year gap between Nemotron-4 (2024) and Nemotron 3 (2026) suggests Nvidia’s open-model cadence is closer to annual-to-biennial than the rapid-fire release schedule seen at labs like Moonshot or Zhipu.
- Enterprise adoption of Nemotron 3 Ultra will hinge on NIM support and enterprise contracts more than raw benchmark scores. Given the roughly 13-point gap to Claude Opus 4.8 on the Artificial Analysis Index, Nvidia’s sales pitch is unlikely to be “smartest model”; it will be “best supported model on the hardware you already run.”
- Efficiency-focused releases will keep gaining relative to pure parameter-count scaling. Kimi K2.7 Code’s reported 30% cut in thinking tokens and Nemotron 3 Ultra’s roughly 10% activation ratio both point toward the same trend: labs competing on cost-per-task, not just headline parameter counts.
- The open-closed benchmark gap narrows but doesn’t close in 2026. Expect the next generation of open-weight flagships, including a likely Nemotron refresh, to close several points of the current 13-point Artificial Analysis Index gap to closed frontier leaders, without fully catching up before the year ends.
These are analyst-style projections based on current release patterns and public statements, not confirmed Nvidia roadmap items.
Frequently Asked Questions
What is Nvidia Nemotron 3 Ultra?
Nemotron 3 Ultra is a 550-billion-parameter open-weight language model from Nvidia, released June 4, 2026. It uses a hybrid Mamba-Attention Mixture-of-Experts architecture with about 55 billion active parameters per token and a 1-million-token context window.
Is Nemotron 3 Ultra free to use?
The model weights are free to download from Hugging Face under Nvidia’s permissive OpenMDW-1.1 license. Running it still requires either your own GPU infrastructure or a paid hosting option through services like OpenRouter or Nvidia’s NIM.
How does Nemotron 3 Ultra compare to Claude Opus 4.8?
Claude Opus 4.8 scores 61.4 on the Artificial Analysis Intelligence Index, versus 47.7-48.2 for Nemotron 3 Ultra, a roughly 13-point gap. Nemotron 3 Ultra is more competitive on specific coding and long-context benchmarks but trails on general frontier reasoning tests like Humanity’s Last Exam.
What is the Nemotron 3 Ultra context window?
Nemotron 3 Ultra supports up to 1 million tokens of context, and Nvidia reports a 94.7% score on the RULER long-context retrieval benchmark at that full context length.
Why did Nvidia release an open-weight model instead of keeping it closed?
Nvidia has not published an explicit line-item rationale, but the strategy is consistent with pulling developers into its software and hardware stack. CEO Jensen Huang has publicly argued that open models “strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty,” a position tied to Nvidia’s broader push against new restrictions on open-weight AI.
What is the Open Weights and American AI Leadership letter?
It’s an industry letter published July 24, 2026, organized with Microsoft and backed by Nvidia, that urges US policymakers not to impose premature restrictions on open-weight AI models. It grew from roughly 25 to about 50 signatories within two days, with OpenAI and Google reportedly added after the initial launch.
Where can developers access Nemotron 3 Ultra?
The model is available for direct download on Hugging Face, through Nvidia’s own build.nvidia.com model catalog and NIM microservices, and via third-party routers like OpenRouter for pay-per-token access without self-hosting.
What other open-weight models launched around the same time?
Nemotron 3 Ultra shipped in the middle of a busy stretch of open-weight releases: MiniMax M3 on June 1, Kimi K2.7 Code from Moonshot AI on June 12, and GLM-5.2 from Zhipu AI on June 13, all within a two-week window.
