跳轉到主要內容

Llama 3.1 Nemotron Ultra 253B v1 (Reasoning)

NVIDIALlama開源權重Llama 3.1 Community License

描述

A 253B parameter derivative of Meta Llama 3.1 405B Instruct, developed by NVIDIA using Neural Architecture Search (NAS) and vertical compression. It underwent multi-phase post-training (SFT for Math, Code, Reasoning, Chat, Tool Calling; RL with GRPO) to enhance reasoning and instruction-following. Optimized for accuracy/efficiency tradeoff on NVIDIA GPUs. Supports 128k context.

發布日期
2025-04-07
參數規模
253.0B
上下文長度
128K
支援模態
text

能力雷達圖

31
general
58
coding
72
reasoning
46
science
70
agents
0
multimodal

排行榜排名

領域#排名分數來源
程式碼能力榜360
27.0
AA
通用能力榜373
33.0
AA
科學能力250
48.0
AA

基準測試分數 (LLM Stats)

(LLM Stats (zeroeval))

Biology

GPQANYU + Cohere + Anthropic (2023)76.0%自報

Code

LiveCodeBench66.3%自報

General

IFEvalGoogle Research (2023)89.5%自報
BFCL v274.1%自報

Math

MATH-50097.0%自報
AIME 202572.5%自報

AA 評測指數

(Artificial Analysis)
Math Index(Artificial Analysis)
63.7
Intelligence Index(Artificial Analysis)
8.9
Math 500(OpenAI (2024), subset of Hendrycks et al. MATH (2021))
1.0
Mmlu Pro(TIGER-Lab (Univ. of Waterloo, Toronto, CMU, 2024))
0.8
Aime(MAA (Mathematical Association of America))
0.7
Gpqa(NYU + Cohere + Anthropic (2023))
0.7
Livecodebench(UC Berkeley + MIT + Cornell (2024))
0.6
Aime 25(MAA (Mathematical Association of America))
0.6
Ifbench(Google Research (2023))
0.4
Scicode(UIUC + Argonne National Lab (2024))
0.3
Tau2(Sierra + U Toronto + Vector Institute (2025))
0.1
Lcr(Artificial Analysis)
0.1
Hle(Center for AI Safety + Scale AI (2025))
0.1
Terminalbench Hard(Stanford × Laude Institute (2026))
0.0

LLM Stats 分類評分

(LLM Stats (zeroeval))
Instruction Following
90
Structured Output
90
Math
80
Physics
80
Reasoning
80
General
80
Biology
80
Chemistry
80
Code
70
Tool Calling
70

定價

輸入價格$0.6 / 1M tokens
輸出價格$1.8 / 1M tokens
混合價格(3:1)$0.9 / 1M tokens

速度

Tokens/秒53.1
首Token延遲0.70s
首回答延遲38.37s

供應商價格排行

供應商價格排行

1 個供應商

供應商輸入輸出
1NVIDIA主要
$0.6
$1.8

比較該模型在不同 API 供應商之間的定價。

外部連結