Sonar Reasoning：完整基准性能指南

other sonar-reasoning 2024-12-18 Proprietary

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

Key Specifications

Vendorother
Versionsonar-reasoning
Release Date2024-12-18
Context Window127072 tokens
Input Modalitiestext
Output Modalitiestext
LicenseProprietary
Documentationhttps://huggingface.co/models

Benchmark Performance

BenchmarkScoreUnitEvaluated AtNotesSource
MMLU87.7%2024-12-185-shotview
HUMANEVAL81.1pass@12024-12-18—view
GSM8K89.8%2024-12-180-shot CoTview
MATH50.7%2024-12-180-shot CoTview
BBH83.4%2024-12-183-shot CoTview
GPQA50.1%2024-12-180-shotview
IFEVAL83.1%2024-12-18prompt_strictview
ARC96.2%2024-12-18challengeview
MUSR69.4%2024-12-180-shotview
WINOGRANDE87.2%2024-12-180-shotview

Pricing

TierPriceCurrency
Input$2 / MtokUSD
Output$8 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://huggingface.co/models · as of 2024-12-18

Compliance

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

Sonar Reasoning

模型概述

Perplexity Sonar Reasoning 推理型在线 RAG 模型, 127K 上下文, 基于 DeepSeek R1 微调, 链式思维推理。

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
Othersonar-reasoning2024-12-18127KtexttextProprietary

基准测试表现

基准得分单位备注
MMLU (Massive Multitask Language Understanding)87.7%5-shot
HumanEval81.1pass@1—
GSM8K (Grade School Math 8K)89.8%0-shot CoT
MATH50.7%0-shot CoT
BBH (BIG-Bench Hard)83.4%3-shot CoT
GPQA50.1%0-shot
IFEval83.1%prompt_strict
ARC96.2%challenge
MUSR69.4%0-shot
WinoGrande87.2%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 87.7，知识推理能力强。
  • HumanEval 81.1，代码生成能力出色。
  • GSM8K 89.8，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