Up to date
nvidia NVIDIA Open Model License (nvidia-open-model-license) out 7 months ago
31.6B parameters, 3B active (6 of 128 experts per token), text in, text out, 256K tokens of context.
Verified 2026-10-25 Rank #629 in LLMs · score 73.4 method v1 Sources 3 Downloads 30d 8.8k · #515 of 619 tracked
Changed this week: 2 new quantized builds on the Hub (37 in all).2026-10-24
Picked by rule from 25 tracked pieces: the lab's own post, the most shared, a field report, the latest.
NVIDIA released Nemotron 3 Ultra on June 4, 2026, as a 550 billion parameter hybrid Mamba-Attention Mixture-of-Experts model designed for long agentic runs.
The 1 pick above, plus these 24. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.
Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.
No API provider serves it, so its downloads are people running it themselves: 37 quantized builds and 3 Spaces.
Top Spaces running it: build-small-hackathon/venue-manager-agent ♥ 1
Weights only, an estimate: leave room for the KV cache. Method
About 19 GB at INT4, 38 GB at FP8, 76 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.
Estimate params × bytes per param × 1.2 · how we estimate
Weights, every dated article, and repo commits on one rail.
Teal is the weights (2026-03-18), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.
59 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Everything else, folded: the full tables, every source.
Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.
| Pipeline tag | text-generation |
|---|---|
| Library | transformers |
| Architectures | NemotronHForCausalLM |
| Model type | nemotron_h |
| Dtype | bfloat16 |
| Params by dtype | BF16: 31,577,937,344 · F32: 2,944 |
| Size band | L |
| MoE | yes |
| Experts | 128 |
| Experts per token | 6 |
| Context length | 262,144 (config.json max_position_embeddings) |
| Languages | en |
| Papers | arXiv:2603.19220 |
| License | NVIDIA Open Model License (nvidia-open-model-license) (other) |
| Hub snapshot | 2026-10-25 |
vLLM: supported · NemotronHForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · NemotronHForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed
vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.
Weights-only memory for Nemotron-Cascade-2-30B-A3B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 76 GB · Single high-end GPU (<80 GB) |
|---|---|
| FP8 (1 byte/param) | 38 GB · Single high-end GPU (<80 GB) |
| INT4 (0.5 bytes/param) | 19 GB · Single GPU (<24 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.
quantized: 37finetune: 20adapter: 2
| Date | Commit | SHA |
|---|---|---|
| 2026-07-09 | Update README.md | 6327cdb |
| 2026-07-09 | Update README.md | 7059384 |
| 2026-07-09 | Update README.md | 895f38b |
| 2026-07-09 | Update README.md | a08121b |
| 2026-05-01 | Update README.md | 8d6a9c7 |
| 2026-04-09 | Update README.md | 5df3d31 |
| 2026-04-09 | Update README.md | 407f169 |
| 2026-03-24 | Update config.json | cfec477 |
| 2026-03-23 | Update README.md | c48c941 |
| 2026-03-23 | Update README.md | 61966c4 |
| 2026-03-21 | Update README.md | faaa1ca |
| 2026-03-20 | Upload nemotron-cascade-2-results.png | 6cac6f8 |
| 2026-03-20 | Update README.md | 9324e20 |
| 2026-03-20 | Update README.md | e8531b3 |
| 2026-03-20 | Upload Nemotron_Cascade_2.pdf | ba96b6d |
| 2026-03-20 | Delete ioi2025 | 9854704 |
| 2026-03-20 | Delete imoproofbench | d0087b7 |
| 2026-03-20 | Delete imo2025 | 7208067 |
| 2026-03-20 | Delete icpcwf2025 | a05e35b |
| 2026-03-20 | Upload folder using huggingface_hub | 34e5197 |
| 2026-03-20 | Update README.md | 0f22793 |
| 2026-03-20 | Upload folder using huggingface_hub | 595c358 |
| 2026-03-20 | Update config.json | 2bdea4a |
| 2026-03-19 | Upload fig/nemotron-cascade-2-results.png | 7d2ba01 |
| 2026-03-19 | Update README.md | a33e221 |
| 2026-03-18 | Upload folder using huggingface_hub | 4a7a2dc |
| 2026-03-18 | initial commit | 0224ac3 |
Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:41 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:13 UTC · Last verified 2026-10-25
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.
