Gemini 1.0 Pro：完整基准性能指南

google 1.0-pro 2023-12-06 Proprietary

深入分析 Gemini 1.0 Pro 在 MMLU、HumanEval、GSM8K、MATH、BBH 等基准上的表现，包含定价与适用场景。

Key Specifications

Vendorgoogle
Version1.0-pro
Release Date2023-12-06
Context Window32768 tokens
Input Modalitiestext
Output Modalitiestext
LicenseProprietary
Documentationhttps://ai.google.dev/gemini-api/docs

Benchmark Performance

BenchmarkScoreUnitEvaluated AtNotesSource
MMLU84.5%2023-12-065-shotview
HUMANEVAL76.4pass@12023-12-06—view
GSM8K82.4%2023-12-060-shot CoTview
MATH58.2%2023-12-060-shot CoTview
BBH79.1%2023-12-063-shot CoTview
GPQA47.7%2023-12-060-shotview
IFEVAL75.1%2023-12-06prompt_strictview
ARC94.1%2023-12-06challengeview
MUSR57.8%2023-12-060-shotview
WINOGRANDE84.5%2023-12-060-shotview

Pricing

TierPriceCurrency
Input$0.5 / MtokUSD
Output$1.5 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://ai.google.dev/pricing · as of 2023-12-06

Compliance

  • Data Residency: US
  • SOC2: ✓
  • HIPAA: ✗
  • GDPR: ✓
  • ISO 27001: ✓

Gemini 1.0 Pro

模型概述

Google Gemini 1.0 Pro 首代模型, 32K 上下文, 在 MMLU/GSM8K 上超越 GPT-3.5, 多模态能力初步集成。

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
Google1.0-pro2023-12-0632KtexttextProprietary

基准测试表现

基准得分单位备注
MMLU (Massive Multitask Language Understanding)84.5%5-shot
HumanEval76.4pass@1—
GSM8K (Grade School Math 8K)82.4%0-shot CoT
MATH58.2%0-shot CoT
BBH (BIG-Bench Hard)79.1%3-shot CoT
GPQA47.7%0-shot
IFEval75.1%prompt_strict
ARC94.1%challenge
MUSR57.8%0-shot
WinoGrande84.5%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 84.5，知识推理能力强。

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