The Information Machine

NVIDIA Nemotron 3 Ultra: Hybrid SSM/MoE Architecture Launch and Benchmarks · history

Version 4

2026-06-07 18:14 UTC · 146 items

What

NVIDIA launched Nemotron 3 Ultra (550B total/55B active parameters, hybrid Mamba SSM/MoE) at Computex 2026 and shipped it June 4, simultaneously announcing the Nemotron Coalition — eight AI labs including Mistral AI, Nous Research, and H Company — with DGX Cloud compute offered to members [7][10]. Independent code benchmarks show near-identical output to GPT-5.5 at ~10x lower cost [16], but the model trails Kimi K2.6 on raw intelligence while offering 3-6x faster inference [17]. Sebastian Raschka independently called the Nemotron 3 Super architecture paper 'the most practically important' of the first half of 2026 [6]. Production reliability concerns persist: one deployment observation found GPU at ~30% utilization with a CPU core bottlenecking requests [21], and CodeRabbit flagged the model as close to baseline in code review with a specific reliability caveat [18].

Why it matters

The cost-performance ratio is compelling for enterprise users, and academic validation of the hybrid SSM/MoE architecture confirms it addresses a real long-context problem rather than being a scale showcase. But unresolved production reliability and the coalition's governance implications — specifically what dependencies DGX Cloud access creates for member labs — are the key variables determining whether the launch converts to durable adoption.

Open questions

  • Does NVIDIA's DGX Cloud subsidy to coalition member labs create dependency relationships that could constrain their independence? The governance terms have not been publicly detailed [10].

  • Does Nemotron 3 Ultra meet production reliability standards? A deployment observation found GPU at ~30% utilization with one CPU core pinned at 100%, suggesting per-request latency is dominated by fixed overhead rather than GPU compute [21], and CodeRabbit flagged reliability as a specific concern [18].

  • Does the ~10x cost advantage over GPT-5.5 hold across task types beyond code generation, and can Nemotron close the gap with Kimi K2.6 on raw intelligence? [16][17]

  • Will new SSM variants (Mamba-3, Gated DeltaNet-2) identified by Raschka as emerging appear in future open-weight models that could compete with or succeed the Nemotron architecture? [6]

Narrative

NVIDIA announced Nemotron 3 Ultra at Jensen Huang's Computex 2026 keynote and shipped it on June 4, 2026. It has 550B total parameters with 55B active — a mixture-of-experts ratio of roughly 10:1 — and uses a hybrid architecture pairing a Mamba-style state-space model (SSM) with a Transformer backbone [1][2]. The SSM handles long sequences without the quadratic memory cost of full attention, making the architecture suited to extended agentic reasoning and multi-step tool use. NVIDIA positioned it as the most intelligent US open-weights model, citing Artificial Analysis's GDPval-AA benchmark in the keynote [3][4]. The release comprises not just the model but a full reproducible reasoning stack: open weights, math SFT training data, and an evaluation framework [5]. Sebastian Raschka, in a June 2026 survey of the year's LLM research, independently named the Nemotron 3 Super architecture paper — which describes the same hybrid SSM/MoE approach — as the most practically important architecture paper of the first half of 2026, citing long-context efficiency as the dominant 2026 architectural priority as models are deployed inside agent harnesses [6].

Alongside the model, NVIDIA formally launched the Nemotron Coalition, a consortium of eight AI labs. Named members include Mistral AI, Nous Research, H Company, Prime Intellect, Black Forest Labs, Cursor AI, LangChain, and Perplexity AI [7][8][9]. NVIDIA offers DGX Cloud compute access to member labs [10]. Mistral AI's CEO Arthur Mensch publicly endorsed the partnership at announcement [11], and Nous Research joined as a new member on launch day [12][13]. The coalition's governance structure — specifically what obligations or dependencies DGX Cloud access creates for member labs — has not been publicly detailed.

Performance comparisons present a mixed picture. Artificial Analysis benchmarks show Nemotron 3 Ultra scoring 47.7 versus Qwen3.6-27B's 45.8 with faster task completion [14][15]. A direct code-generation comparison showed near-identical output to GPT-5.5 at roughly 10x lower cost ($0.051 vs $0.57) [16]. The model reportedly trails Kimi K2.6 on raw intelligence while delivering 3-6x faster inference at lower cost [17]. CodeRabbit found the model close to baseline in code review but flagged production reliability as a specific concern [18][19]. One evaluator cited an agent benchmark score of 89.9% versus Claude Opus 4.7's 93.5% [20]. A deployment observation flagged GPU at ~30% utilization with one CPU core pinned at 100%, suggesting per-request latency is dominated by a large fixed cost rather than GPU computation [21].

The coalition drew ideological criticism from SemiAnalysis, which on June 5 called it 'communist committee-style' and argued that free-market competition — including using Chinese open models like Kimi — is the correct path for American open-source AI [22][23]. This framing inverts the expected US-China narrative: SemiAnalysis describes an American industry consortium as the non-market option and a Chinese model as the free-market one. Infomly separately confirmed that Kimi K2.6 does lead Nemotron on raw intelligence [17]. Grok/xAI stated it would not adopt Nemotron 3 Ultra, consistent with xAI's approach of building its own models end-to-end [24]. Enterprise adoption has been broad: Glean, Nebius, Simplismart, FriendLI, and Kilocode all announced same-day or near-day support [25][26][27], and the model was added to HuggingChat on June 5 [28].

Timeline

  • 2026-06-01: Rohan Paul announces Nemotron 3 Ultra will ship within days, summarizing its hybrid SSM/MoE architecture and long-context advantages. [2]
  • 2026-06-01: Skeptical commentary emerges: Nemotron models historically rank ~5th in open-source rankings and 550B scale is called a compute showcase. [36][37]
  • 2026-06-02: Mistral AI CEO Arthur Mensch posts support for joining the Nemotron Coalition. [11]
  • 2026-06-03: Artificial Analysis notes NVIDIA cited its GDPval-AA benchmark in Jensen Huang's Computex 2026 keynote to support performance claims. [4][3]
  • 2026-06-04: Nemotron 3 Ultra officially ships; becomes available on Nebius Token Factory, Kilocode, and FriendLI on day 0. [26][38][39][27]
  • 2026-06-04: NVIDIA formally launches the Nemotron Coalition of eight AI labs, offering DGX Cloud compute access to member labs. [7][8][10]
  • 2026-06-04: Nous Research joins the Nemotron Coalition. [12][13]
  • 2026-06-04: Glean announces same-day enterprise support for Nemotron 3 Ultra. [25][40]
  • 2026-06-04: Artificial Analysis independent benchmarks: Nemotron 3 Ultra scores 47.7 versus Qwen3.6-27B's 45.8 with faster task completion. [15][14]
  • 2026-06-04: CodeRabbit evaluation: Nemotron 3 Ultra lands close to baseline with a reliability caveat for production use. [18][33][19]
  • 2026-06-04: One evaluator cites agent benchmark score of 89.9% for Nemotron 3 Ultra versus 93.5% for Claude Opus 4.7. [20]
  • 2026-06-04: Grok/xAI states it will not adopt Nemotron 3 Ultra; xAI builds and trains its own models end-to-end. [24]
  • 2026-06-05: Side-by-side comparison shows Nemotron 3 Ultra producing near-identical code output to GPT-5.5 at roughly 10x lower cost ($0.051 vs $0.57). [16]
  • 2026-06-05: SemiAnalysis calls the Nemotron Coalition 'communist committee-style' and states it will use Chinese open models (Kimi) instead for open-source work. [22][23]
  • 2026-06-05: Infomly reports Nemotron 3 Ultra trails Kimi K2.6 on raw intelligence but delivers 3-6x faster inference at lower cost. [17]
  • 2026-06-05: Deployment observation: GPU at ~30% utilization, one CPU core pinned at 100%, per-request latency dominated by fixed overhead. [21]
  • 2026-06-05: Nemotron 3 Ultra added to HuggingChat, expanding public access beyond inference API providers. [28]
  • 2026-06-06: Sebastian Raschka names the Nemotron 3 Super architecture paper the most practically important of the first half of 2026, citing long-context efficiency as the dominant 2026 architectural priority. [6]

Perspectives

NVIDIA (official)

Nemotron 3 Ultra is the most intelligent US open-weights model built for long-running agentic workloads; the Nemotron Coalition advances American open-source frontier AI through lab collaboration with DGX Cloud compute support.

Evolution: Consistent with NVIDIA's Nemotron positioning; the coalition announcement added an organizational dimension to what was previously a single-model story.

[29][30][7][10][31]

Sebastian Raschka (@rasbt)

The Nemotron 3 Super architecture paper — describing the same hybrid SSM/MoE approach used in Ultra — is the most practically important architecture paper of the first half of 2026; long-context efficiency via SSM is the correct architectural priority as models enter agent harnesses.

Evolution: New voice this pass; provides independent academic validation of the architectural rationale from a researcher with no coalition affiliation.

[6]

Artificial Analysis (@ArtificialAnlys)

Independent benchmarks confirm Nemotron 3 Ultra edges Qwen3.6-27B (47.7 vs 45.8) with faster task completion; their GDPval-AA metric was cited in NVIDIA's keynote.

Evolution: Consistent empirical testing role; elevated to keynote prominence by NVIDIA's citation.

[15][4][3][32]

Rohan Paul (@rohanpaul_ai)

Positive on architectural rationale and cost case: SSM enables long-context processing without attention bottlenecks, and the ~10x cost gap versus GPT-5.5 on code tasks is compelling for developers.

Evolution: Consistent; provided pre-release architecture summary and followed up with a direct GPT-5.5 price comparison on launch day.

[2][16]

CodeRabbit (@coderabbitai)

Measured: Nemotron 3 Ultra performs close to baseline in code review benchmarks but production reliability is a specific concern for production workloads.

Evolution: Consistent; full blog post confirms the reliability caveat raised at launch.

[18][33][19][34]

SemiAnalysis (@SemiAnalysis_)

The Nemotron Coalition is 'communist committee-style' and the wrong model for American open-source AI; free-market competition including use of Chinese open models like Kimi is preferable.

Evolution: Consistent since June 5; coalition membership details reinforce rather than change their critique.

[22][23]

Mistral AI and coalition members

Supportive of open frontier model collaboration with NVIDIA; Mistral's CEO expressed enthusiasm for building frontier open-source AI together with NVIDIA as a coalition member.

Evolution: Coalition membership formalized and publicly endorsed at launch; no subsequent statements modifying this position.

[35][11][12][13]

Skeptics and comparators (ECLresearch, Aero, Infomly)

Nemotron 3 Ultra is primarily a speed and cost showcase; Nemotron models historically rank ~5th in open-source, and Kimi K2.6 leads on raw intelligence even as Nemotron is faster and cheaper.

Evolution: Infomly's Kimi K2.6 comparison added a specific capability counter-reference; position otherwise consistent.

[36][37][17]

Tensions

  • NVIDIA claims Nemotron 3 Ultra is the most intelligent US open-weights model; skeptics argue Nemotron historically ranks ~5th in open-source and Kimi K2.6 leads on raw intelligence, even as Nemotron offers faster inference at lower cost. [29][36][17]
  • NVIDIA's keynote cited GDPval-AA to support performance claims; CodeRabbit's code review benchmark and an agent benchmark score (89.9% vs Claude Opus 4.7's 93.5%) show more modest results than the headline framing suggests. [4][18][20]
  • A code-generation comparison shows ~10x cost savings versus GPT-5.5, but a deployment observation found CPU-bottlenecked latency and CodeRabbit flagged production reliability, qualifying the cost-advantage case for production use. [16][21][18]
  • SemiAnalysis argues the Nemotron Coalition's coordinated structure contradicts American free-market principles; NVIDIA frames the coalition as advancing American open-source AI through DGX Cloud support and lab collaboration. [22][23][7][10]
  • SemiAnalysis prefers Chinese open models (Kimi) over an American-organized coalition, labeling the consortium as the non-market option; Infomly separately confirms Kimi K2.6 does lead Nemotron on raw intelligence. [23][17]

Sources

  1. [1] NVIDIA has announced Nemotron 3 Ultra at Jensen Huang's Computex 2026 keynote. At 550 billion total parameters (55 billi... — reactive:nvidia-nemotron-ultra (2026-06-04)
  2. [2] Nemotron 3 Ultra will be available from Nvidia in few days. — Rohan Paul Twitter (2026-06-01)
  3. [3] Jensen Huang’s keynote at Computex used Artificial Analysis benchmarks to communicate the performance of Nemotron 3 Ultr... — reactive:nvidia-nemotron-ultra (2026-06-03)
  4. [4] NVIDIA used the GDPval-AA benchmark to promote the launch of NVIDIA Nemotron 3 Ultra https://t.co/N6e4ItzYev — reactive:nvidia-nemotron-ultra (2026-06-03)
  5. [5] NVIDIA is not merely releasing Nemotron 3 Ultra. It is releasing a reproducible reasoning stack: open weights, math SFT ... — reactive:nvidia-nemotron-ultra (2026-06-05)
  6. [6] LLM Research Papers: The 2026 List (January to May) — Ahead of AI (2026-06-06)
  7. [7] NVIDIA Launches Nemotron Coalition of Leading Global AI Labs to Advance Open Frontier Models | NVIDIA Newsroom — reactive:open-model-capability-gap
  8. [8] Nvidia's Nemotron coalition brings eight AI labs together to build open frontier models | Tom's Hardware — reactive:open-model-capability-gap
  9. [9] @NVIDIAAI @hcompany_ai @NousResearch @PrimeIntellect @bfl_ai @cursor_ai @LangChain @MistralAI @perplexity_ai @Reflection... — reactive:nvidia-nemotron-ultra (2026-06-05)
  10. [10] Nvidia offers DGX Cloud to AI foundation model labs for new ... — reactive:nvidia-nemotron-ultra
  11. [11] RT @arthurmensch: Looking forward to building frontier open source AI models together with @Nvidia as we join the Nemotr... — reactive:nvidia-nemotron-ultra (2026-06-02)
  12. [12] Nous Research (AI lab open-source) baru join coalition Nvidia buat model Nemotron. — reactive:nvidia-nemotron-ultra (2026-06-05)
  13. [13] Excited to see @NousResearch join NVIDIA's Nemotron Coalition! — reactive:nvidia-nemotron-ultra (2026-06-04)
  14. [14] Not only did Nemotron 3 Ultra beat Qwen3.6-27b in raw AA benchmark (47.7 vs 45.8) but it also completed the tasks faster... — reactive:nvidia-nemotron-ultra (2026-06-04)
  15. [15] Nemotron 3 Ultra was launched today, including a focus on low latency agentic performance. We tested it against peers un... — reactive:nvidia-nemotron-ultra (2026-06-04)
  16. [16] Nemotron 3 Ultra vs GPT-5.5 on atomic[.]chat, a desktop app that runs LLMs locally. — Rohan Paul Twitter (2026-06-05)
  17. [17] NVIDIA Nemotron 3 Ultra Trails Kimi K2.6 on raw intelligence but delivers 3-6x faster inference at lower cost. — reactive:nvidia-nemotron-ultra (2026-06-05)
  18. [18] In our CodeRabbit benchmark, Nemotron 3 Ultra landed close to the baseline: — reactive:nvidia-nemotron-ultra (2026-06-04)
  19. [19] Nemotron 3 Ultra makes the case for fast, open coding models — reactive:nvidia-nemotron-ultra
  20. [20] @turingou 我一直堅信固定工作流的agent不需要SoTA，英偉達nemotron 3 ultra在agent benchmark 拿了89.9%，OPUS 4.7拿了93.5%。 — reactive:nvidia-nemotron-ultra (2026-06-04)
  21. [21] GPU at ~30% utilization. One CPU core pinned at 100%. Per-request latency dominated by a large fixed cost. — reactive:nvidia-nemotron-ultra (2026-06-05)
  22. [22] The path for American open models to catch up to Chinese open models is not through a “coalition,” but through the found… — SemiAnalysis Twitter (2026-06-05)
  23. [23] We fundamentally disagree with the communist committee-style “Nemotron Coalition” approach to developing OSS models, and… — SemiAnalysis Twitter (2026-06-05)
  24. [24] @xiaosun86 @NVIDIAAI No, Grok isn't adopting Nemotron 3 Ultra. xAI builds and trains its own models end-to-end. Strong b... — reactive:nvidia-nemotron-ultra (2026-06-01)
  25. [25] 🎉 We’re announcing support for @nvidia Nemotron Ultra 3! — reactive:nvidia-nemotron-ultra (2026-06-04)
  26. [26] NVIDIA Nemotron™ 3 Ultra is now live on Nebius Token Factory. — reactive:nvidia-nemotron-ultra (2026-06-04)
  27. [27] Run NVIDIA's Most Powerful Open Reasoning Model on Day 0 — reactive:nvidia-nemotron-ultra
  28. [28] Nemotron 3 Ultra is on HuggingChat ☘️ — reactive:nvidia-nemotron-ultra (2026-06-05)
  29. [29] Nemotron 3 Ultra 550B (A55B) is the largest Nemotron 3 model to date, and the most intelligent US open weights model rig... — reactive:nvidia-nemotron-ultra (2026-06-01)
  30. [30] NVIDIA ships Nemotron 3 Ultra, a 550B MoE model for long-running agents. It runs 5x faster at inference and reduces cost... — reactive:nvidia-nemotron-ultra (2026-06-04)
  31. [31] NVIDIA Nemotron 3 Ultra Powers Faster, More Efficient Reasoning ... — reactive:nvidia-nemotron-ultra
  32. [32] Nemotron 3 Ultra - Intelligence, Performance & Price Analysis — reactive:nvidia-nemotron-ultra
  33. [33] The caveat: reliability. — reactive:nvidia-nemotron-ultra (2026-06-04)
  34. [34] Nemotron 3 Ultra makes the case for fast, open coding models — reactive:nvidia-nemotron-ultra
  35. [35] Mistral AI partners with NVIDIA to accelerate open frontier models — reactive:open-model-capability-gap
  36. [36] @jun_song Solid skepticism. Nemotron 3 Ultra at 550B is a compute flex, but the open-weight ecosystem moved past "just b... — reactive:nvidia-nemotron-ultra (2026-06-01)
  37. [37] @scaling01 Nemotron models are always released to be #5 on open-source while using the latest nvidia tricks to be fast a... — reactive:nvidia-nemotron-ultra (2026-06-01)
  38. [38] Nemotron 3 ultra, Qwen 3.7 plus, laguna m.1, step 3.7 flash ALL FREE on kilocode. but which to use? BENCHMARK TEST TIME.... — reactive:nvidia-nemotron-ultra (2026-06-04)
  39. [39] 550 billion parameters. Open weights. Ships today. — reactive:nvidia-nemotron-ultra (2026-06-04)
  40. [40] Excited to expand model choice in @Glean with @NVIDIA Nemotron 3 Ultra. — reactive:nvidia-nemotron-ultra (2026-06-04)