Back to Qwen
Alibaba / Alibaba Cloud / Qwen Team /

Qwen 2.5 72B Instruct
Qwen 2.5 72B InstructFeatured

Open WeightsChat & ReasoningtextUpdated September 19, 2024

Qwen2.5-72B-Instruct: Alibaba's Flagship Open LLM

Model Overview

Qwen2.5-72B-Instruct is Alibaba's flagship open-source large language model, featuring 72 billion parameters. Released as part of the Qwen2.5 model family in September 2024, it achieves top-tier performance across instruction following, coding, mathematics, and reasoning benchmarks. Qwen2.5-72B-Instruct supports a 128K token context window, multiple languages, and structured output formats including JSON mode. It consistently ranks at or near the top of open-model leaderboards.


✨ Key Features

Specification Table
FeatureDescription
72B ParametersAlibaba's largest openly released LLM, competitive with frontier models
128K ContextExtended context for long-form documents and multi-turn conversations
MultilingualStrong performance across 29 languages including Chinese, English, Arabic, French, German
Structured OutputNative JSON mode and structured generation support
Top Leaderboard PerformanceConsistently ranks among best open-source models on AlpacaEval and Open LLM Leaderboard

📊 Benchmarks

Specification Table
BenchmarkQwen2.5-72B-Instruct
MMLU86.1%
MATH83.1%
HumanEval86.6%
LiveCodeBench65.9%

🔗 Resources

Specification Table
ResourceLink
HuggingFaceQwen/Qwen2.5-72B-Instruct
GitHubQwenLM/Qwen2.5

📜 License

Qwen License — Free for commercial use for deployments serving fewer than 100 million monthly active users.

Key Features

Massive Pre-training Scale: 72.7 billion parameters trained on 18 trillion high-quality tokens

Feature 01

Long Context Support: Native support for input contexts up to 131,072 (128K) tokens and generation up to 8,192 tokens

Feature 02

Multilingual Versatility: Full comprehension and response capabilities across 29+ languages

Feature 03

Frontier Code & Math Reasoning: SOTA open-weights performance on competitive coding, math, Program-of-Thought, and tool-integrated reasoning

Feature 04

Structured Data & Instruction Alignment: Specialized fine-tuning for structured output generation (JSON, tables) and strict system instruction adherence

Feature 05

You might also want to compare

Verified Sources

Tags

llmqwenalibabaopen-weights128k-contextcodingmath

Model Specs

open-weights

Parameters

72.7B

Context Window

128K tokens

License

Other/Custom

Deployment

self-hostableapi-only

Resources & Links

Lineage

Curator Notes

Verified paper arXiv:2412.15115 and open-weights release from Alibaba Qwen Team.

Compare Specs

Compare parameters, context windows, modalities, and benchmark scores of this model side-by-side with others.

Compare Model