Phi-3.5 MoE: Complete Benchmark Performance Guide

other phi-3-5-moe 2024-08-16 MIT

In-depth analysis of Phi-3.5 MoE performance on MMLU, HumanEval, GSM8K, MATH, BBH benchmarks, including pricing and use cases.

Key Specifications

Vendorother
Versionphi-3-5-moe
Release Date2024-08-16
Context Window128000 tokens
Input Modalitiestext
Output Modalitiestext
LicenseMIT
Documentationhttps://huggingface.co/models

Benchmark Performance

Benchmark Score Unit Evaluated At Notes Source
MMLU 75.7 % 2024-08-16 5-shot view
HUMANEVAL 69.5 pass@1 2024-08-16 — view
GSM8K 75.5 % 2024-08-16 0-shot CoT view
MATH 38.2 % 2024-08-16 0-shot CoT view
BBH 79.4 % 2024-08-16 3-shot CoT view
GPQA 40.2 % 2024-08-16 0-shot view
IFEVAL 74.3 % 2024-08-16 prompt_strict view
ARC 91.9 % 2024-08-16 challenge view
MUSR 50.3 % 2024-08-16 0-shot view
WINOGRANDE 78.3 % 2024-08-16 0-shot view

Pricing

Tier Price Currency
Input$0.2 / MtokUSD
Output$0.2 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://huggingface.co/models · as of 2024-08-16

Compliance

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

Phi-3.5 MoE

Model Overview

Microsoft Phi-3.5 MoE 42B 混合专家模型, 128K 上下文, 总参 42B/活跃 6.6B, 性能接近 70B 稠密模型。

Core Specifications

Vendor Version Release Date Context Window Input Modalities Output Modalities License
Other phi-3-5-moe 2024-08-16 128K text text MIT

Benchmark Performance

Benchmark Score Unit Notes
MMLU (Massive Multitask Language Understanding) 75.7 % 5-shot
HumanEval 69.5 pass@1 —
GSM8K (Grade School Math 8K) 75.5 % 0-shot CoT
MATH 38.2 % 0-shot CoT
BBH (BIG-Bench Hard) 79.4 % 3-shot CoT
GPQA 40.2 % 0-shot
IFEval 74.3 % prompt_strict
ARC 91.9 % challenge
MUSR 50.3 % 0-shot
WinoGrande 78.3 % 0-shot

Pricing

Input Output Cache Read Cache Write
— — — —

per million tokens

Strengths

  • Mixture-of-Experts architecture.

Weaknesses

  • Proprietary, not self-hostable.

Use Cases

  • General chat and Q&A

References