Llama Nemotron Ultra 253BOpen Weight

Released

Llama Nemotron Ultra 253B is an AI model released by NVIDIA on Tuesday, Apr 8 2025, 21 days after Llama Nemotron (Nano/Super). It is an open-weight model — the trained weights are available to download and run. It is a 253B parameter model. Benchmark results (shown below) cover LiveCodeBench, and GPQA Diamond.

Benchmarks

Coding

LiveCodeBench
66.3%
#12 of 12

Reasoning & science

GPQA Diamond
76%
#37 of 57

About

Llama Nemotron Ultra 253B, released April 8, 2025, completed the reasoning family NVIDIA had announced at GTC three weeks earlier. It was compressed by neural architecture search from Meta's Llama 3.1 405B down to 253 billion parameters — sized deliberately so the strongest open reasoning model of the family fit a single 8-GPU H100 node — and post-trained for switchable step-by-step reasoning.

At launch NVIDIA reported 76.0% on GPQA Diamond and 66.3% on LiveCodeBench, competitive with DeepSeek-R1 — the model that had defined the open reasoning tier — at just over a third of R1's total parameter count. The weights and much of the post-training data went out openly, and Ultra stood as NVIDIA's open flagship until the in-house Nemotron 3 generation began that December.

Compare Llama Nemotron Ultra 253B with

Llama Nemotron Ultra 253B

Suggested comparisons

Frequently asked questions

Llama Nemotron Ultra 253B was released by NVIDIA on Tuesday, Apr 8 2025.

Llama Nemotron Ultra 253B was built by NVIDIA. The dominant AI chipmaker, publishing the open Nemotron model family. Ships weights, datasets and training recipes to seed demand for its hardware rather than to sell a chatbot.

Llama Nemotron Ultra 253B reports 2 tracked benchmark scores — LiveCodeBench: 66.3%; GPQA Diamond: 76%. Scores are the figures published at release by NVIDIA.

Llama Nemotron Ultra 253B is reported at 253B parameters.

Partly. Llama Nemotron Ultra 253B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense.

NVIDIA's previous tracked release was Llama Nemotron (Nano/Super) on Mar 18 2025, 21 days earlier. It was followed by Nemotron Nano 2 on Aug 18 2025.

All NVIDIA releases

9 tracked