All news

NVIDIA announces Nemotron 3 Ultra — 500B/50B open MoE, weights expected this week

Jensen Huang unveiled Nemotron 3 at Computex on 2026-06-01. The Ultra tier is 500B/50B-active MoE, NVFP4-trained on Blackwell. AA scores it Intelligence Index 48 — top of US open-weights, ahead of Gemma 4 31B, Nemotron 3 Super, gpt-oss-120B. Weights not yet on HF; this is announcement-only.

NVIDIA CEO Jensen Huang unveiled the Nemotron 3 family at Computex 2026 on 2026-06-01 (NVIDIA newsroom). The flagship Nemotron 3 Ultra is a 500B-parameter mixture-of-experts model with 50B active per token, trained in NVFP4 on Blackwell. Weights are not yet downloadable — the canonical Hugging Face URL nvidia/NVIDIA-Nemotron-3-Ultra returns HTTP 401 as of this writing — but Artificial Analysis published independent benchmarks under embargo and NVIDIA put the model into NIM preview, so the spec sheet is public even if the .safetensors aren’t.

What’s been confirmed

From NVIDIA’s newsroom announcement:

Family memberParams (total)Active per tokenContextStatus
Nemotron 3 Nano30 Bup to 3 B1 M tokensReleased 2025-12-15
Nemotron 3 Super~100 Bup to 10 B(not stated)“First half of 2026”
Nemotron 3 Ultra~500 Bup to 50 B(not stated)“First half of 2026”

All three are hybrid-MoE; Super and Ultra were trained in NVFP4 format on Blackwell — the same low-precision format we benchmarked separately on DGX Spark this week with the Qwen 3.6 35B-A3B-NVFP4 checkpoint. Training-data release is unusually generous: NVIDIA published 3 trillion tokens across pretraining, post-training, RL, and a dedicated “Nemotron Agentic Safety Dataset” — that’s not standard practice for vendor open-weights releases.

What Artificial Analysis is reporting

Per the AA launch article, Ultra hits Intelligence Index 48 — top of the US open-weights leaderboard:

“This is well ahead of the next strongest US open weights models, Gemma 4 31B (39), Nemotron 3 Super (36) and gpt-oss-120b (33).”

NVIDIA additionally claims 300+ output tokens/second, up to 5× faster inference, and ~30 % lower cost versus the leading alternatives (Cryptobriefing recap). Treat the speed and cost numbers as vendor self-report until a public leaderboard with reproducible methodology lands them — AA’s Intelligence Index 48 is the only third-party-verified score for now.

For our roster, Nemotron 3 Ultra would slot between DeepSeek V4-Pro (1.6 T) and Kimi K2.6 (1 T) at the frontier end, sitting an order of magnitude above the existing Nemotron 3 Super 120B-A12B entry — note that NVIDIA’s newsroom uses hedged round numbers (“approximately 100 B / up to 10 B” for Super, “about 500 B / up to 50 B” for Ultra), while actual ship configs come in slightly higher: Super shipped at 120B/12B per its HF card, so Ultra is plausibly closer to ~550B/55B at release.

Where it’ll be runnable

Per NVIDIA’s announcement, distribution at release will span:

  • Hugging Face — model card + weights
  • NVIDIA NIM microservice at build.nvidia.com
  • OpenRouter — API tier
  • ModelScope — China mirror
  • Provider rollouts (Baseten, DeepInfra, Fireworks, FriendliAI, Together AI for Nano today; rest of family rolling in)

NVIDIA’s reference target is Blackwell — the model was trained in NVFP4, so a Blackwell GPU with NVFP4 inference kernels (vLLM CUTLASS-FP4 MoE, TRT-LLM) gets the architectural fit. Hopper owners will need to re-quantize to FP8.

What hardware can host it at NVFP4

Working math: 500B params × ~4.5 bits/param NVFP4 ≈ 290 GB of weight footprint on-disk and in VRAM, before any KV cache. Add ~5 GB activations for a 50B-active MoE, plus FP8 KV cache scaling with context length (~5 GB per request at 32 K, ~15 GB at 100 K for a 500B-class model — exact numbers depend on layer count + KV head count, which the announcement didn’t publish, so treat as ±20 %).

NVFP4 is Blackwell-only. Hopper and earlier have to re-quantize to FP8 (which roughly doubles the weight footprint to ~580 GB).

Build (from our catalog)Usable VRAMVerdict at NVFP4
DGX Spark 128 GB119 GB✗ Too small (weights alone don’t fit)
2× DGX Spark240 GB✗ Too small (~50 GB short of weights)
4× DGX Spark at $19,500488 GB✓ Cheapest viable path. Holds weights with ~200 GB headroom — comfortable single-user / 32K context. Caveat: Spark-cluster interconnect (Thunderbolt 5 / EthernetMax) bottlenecks prefill vs NVLink.
8× DGX Spark at $43,500976 GB✓ Comfortable for long-context (100K+) batched serving
RTX PRO 6000 Blackwell 96 GB93 GB✗ Way too small
2× RTX PRO 6000188 GB✗ Too small (~100 GB short)
4× RTX PRO 6000 at $38,000372 GB✓ Tight but works — proper NVLink bandwidth, ~80 GB headroom (good for 32K context, marginal for 100K). The “actually-fast” sweet spot.
8× RTX PRO 6000 at $78,000744 GB✓ Comfortable for 100K-context batched serving
12× RTX PRO 6000 at $118,0001,116 GB✓ Enterprise serving box, lots of room
1× B200 180 GB176 GB✗ Single card too small; 2× B200 (~$94K) would just barely fit weights
8× B200 at $475,0001,404 GB✓ Enterprise, overkill but standard procurement target for NVIDIA-stack shops

Hopper at FP8 (re-quantized):

BuildUsableVerdict at FP8 (~580 GB weights)
H100 80 GB ×8 at $280,000620 GB✓ Tight — ~40 GB headroom, single-user short-context only
H200 141 GB ×8 at $380,0001,100 GB✓ Comfortable at FP8, real headroom for batched / long-context

Not capable for Nemotron 3 Ultra inference: Strix Halo (no NVFP4 hardware, also memory-bound), AMD Instinct (ROCm, no NVFP4 path), Apple Silicon (uses MLX-FP4 separately — different format, would need an MLX-quantized re-export from NVIDIA or the community). Mac Studio M3 Ultra at the maxed 256 GB tier is too small even before considering quant compatibility.

The cheap-vs-fast tradeoff for a single-user agentic workload:

  • $19,500 — 4× DGX Spark cluster: cheapest at-spec config. Spark’s GB10 has full NVFP4 silicon, but the interconnect is the bottleneck — expect single-stream decode in the 30-50 tok/s range and long-context prefill measured in tens of seconds.
  • $38,000 — 4× RTX PRO 6000 Blackwell with NVLink: ~2× the cost, but proper interconnect. Likely 80-120 tok/s single-stream and seconds-not-tens-of-seconds prefill at 32K. The price/performance home for serious agentic work.

The mini-PC / laptop tier today isn’t Nemotron Super. Nemotron 3 Super (120B/12B-A, NVIDIA Open Model License, March 2026) is already in the catalog, and on the benchmarks we track Qwen 3.6 35B-A3B (April, Apache 2.0) beats it across the board: TB2 51.5 vs 31, SWE-V 73.4 vs 60.5, GPQA 86 vs 79.2, HLE 21.4 vs 18.3 — and Qwen is 3.5× smaller in active params and ships under a more permissive license. The interesting question for 128 GB-class hardware isn’t “wait for Super” — it’s whether Qwen 3.6 35B-A3B-NVFP4 (already running ~97 t/s on a single Spark) or a future MiMo / Step variant takes the bucket.

What we’re not saying

  • No tok/s numbers from us yet. We benchmark when weights are live and the recipe is published.
  • No license details on Ultra specifically — Nemotron 3 Nano shipped under the NVIDIA Open Model License; assume the family inherits but confirm against Ultra’s model card when published.
  • No context-window number — NVIDIA’s newsroom quotes 1 M context only for Nano; Super and Ultra context windows weren’t stated.
  • No HF URL. We’ll add the model card link to our dataset and write a follow-up performance post once nvidia/NVIDIA-Nemotron-3-Ultra resolves with a 200.

Sources