Nemotron 3 Ultra is frontier-level, five times faster, 30% cheaper than the best open models, and completely open.

Judged as of 2026-06-01. The claim has four parts about Nemotron 3 Ultra: (1) accuracy at frontier level on major benchmarks, meaning comparable to leading frontier or top open models; (2) about 5x faster, measured as throughput or tokens/s against the comparison models NVIDIA names; (3) about 30% lower cost; (4) completely open, meaning the weights are released and the training data and recipes are published as well. True: NVIDIA's published benchmarks, model card, and independent evaluations such as Artificial Analysis support all four (speed of roughly 4-6x, cost savings of roughly 25-35%). Mostly true: all four broadly hold with a small numeric slip. Mixed: about two of the four hold, for example the speed claim holds only against certain baselines, or the model is open-weight but its data is not open. Mostly false: only one part holds, or several are clearly overstated. False: the model is not open and its benchmarks fall well short of frontier, or the speed and cost figures are contradicted.

Set before the research began.

1. E1 NVIDIA Nemotron 3 Ultra NVIDIA Research (NVIDIA Nemotron lab) · 4 June 2026 Own record

   “Nemotron 3 Ultra achieves 5.9x, 4.8x, and 1.6x higher inference throughput compared to GLM-5.1-754B-A40B, Kimi-K2.6-1T-A32B, and Qwen-3.5-397B-17B respectively on the 8k token input / 64k token output setting.”

2. E2 NVIDIA Nemotron 3 Ultra released: fast, intelligent, and open Artificial Analysis · 4 June 2026

   “Nemotron 3 Ultra scores 47.7 on the Artificial Analysis Intelligence Index, well ahead of the next strongest US open weights models, Gemma 4 31B (39.2), Nemotron 3 Super (36.0) and gpt-oss-120b (33.3), but behind the Chinese-led open weights frontier (Kimi K2.6 at 53.9).”

3. E3 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Hugging Face Hugging Face (model card authored by NVIDIA) · 24 June 2026 Own record

   “Major portions of the pre-training corpus are released in the Nemotron-Pre-Training-Datasets collection.”

4. E4 NVIDIA Nemotron 3 Ultra Leads Open Models on Accuracy and Efficiency in Agentic RTL Coding | NVIDIA Technical Blog NVIDIA Technical Blog · 27 July 2026 Own record

   “the MoE design improves accuracy per active parameter, delivering up to 5X higher throughput and 30% lower costs compared to other open models.”

Artificial Analysis, an independent evaluator, scores Nemotron 3 Ultra at 47.7 on its Intelligence Index. That puts it ahead of every other US open-weight model but behind the leading open model, Kimi K2.6, at 53.9 (E2), so the model is near the frontier but not at the top. NVIDIA's own release page reports throughput gains of 5.9x and 4.8x over GLM-5.1 and Kimi K2.6, but only 1.6x over Qwen-3.5 (E1), so the 5x figure holds only against some of the models NVIDIA names. The 30% cost saving appears only as a general NVIDIA statement against unnamed open models, with no measurement shown (E4). The model card says "major portions" of the pre-training corpus are released (E3), which does not show that all training data is open, so the evidence does not show the model meets the "completely open" standard.