Llama 3.3 70B：完整基准性能指南

meta 3.3-70b 2024-12-06 Llama 3.3 Community License

深入分析 Llama 3.3 70B 在 MMLU、HumanEval、GSM8K、MATH、BBH 等基准上的表现，包含定价与适用场景。

Key Specifications

Vendormeta
Version3.3-70b
Release Date2024-12-06
Context Window128000 tokens
Input Modalitiestext
Output Modalitiestext
LicenseLlama 3.3 Community License
Documentationhttps://llama.meta.com/docs/

Benchmark Performance

BenchmarkScoreUnitEvaluated AtNotesSource
MMLU83.4%2024-12-065-shotview
HUMANEVAL87pass@12024-12-06—view
GSM8K86.9%2024-12-060-shot CoTview
MATH73.8%2024-12-060-shot CoTview
BBH83.9%2024-12-063-shot CoTview
GPQA52.8%2024-12-060-shotview
IFEVAL79.3%2024-12-06prompt_strictview
ARC93.9%2024-12-06challengeview
MUSR62.3%2024-12-060-shotview
WINOGRANDE86.8%2024-12-060-shotview

Pricing

TierPriceCurrency
Input$0.9 / MtokUSD
Output$0.9 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://ai.meta.com/blog/ · as of 2024-12-06

Compliance

  • Data Residency: self-host
  • SOC2: ✗
  • HIPAA: ✗
  • GDPR: ✗
  • ISO 27001: ✗

Llama 3.3 70B

模型概述

Meta Llama 3.3 70B 开源旗舰, 128K 上下文, 性能超越 Llama 3.1 405B, 接近 GPT-4o 水平。

核心规格

厂商版本发布日期上下文窗口输入模态输出模态许可证
Meta3.3-70b2024-12-06128KtexttextLlama 3.3 Community License

基准测试表现

基准得分单位备注
MMLU (Massive Multitask Language Understanding)83.4%5-shot
HumanEval87.0pass@1—
GSM8K (Grade School Math 8K)86.9%0-shot CoT
MATH73.8%0-shot CoT
BBH (BIG-Bench Hard)83.9%3-shot CoT
GPQA52.8%0-shot
IFEval79.3%prompt_strict
ARC93.9%challenge
MUSR62.3%0-shot
WinoGrande86.8%0-shot

定价

输入输出缓存读取缓存写入
————

每百万 token

优势

  • MMLU 得分 83.4，知识推理能力强。
  • HumanEval 87.0，代码生成能力出色。
  • GSM8K 86.9，数学推理稳健。

劣势

  • 闭源专有模型，不支持自托管。

适用场景

  • 代码生成与调试

参考文献