Qwen2

// Overview

Qwen2 is the open-weight large language model family Alibaba's Qwen team released on June 7, 2024, succeeding Qwen1.5. It ships in five sizes: dense Qwen2-0.5B, Qwen2-1.5B, Qwen2-7B and Qwen2-72B, plus a Mixture-of-Experts model, Qwen2-57B-A14B, that activates roughly 14B of its 57B parameters per token via 64 routed and 8 shared experts. Every size uses Grouped Query Attention (GQA), and the family shares a 151,643-token multilingual vocabulary covering English, Chinese and 27 additional languages.

Each base model ships alongside an instruction-tuned -Instruct variant. Context length varies by size: 32K tokens for the 0.5B and 1.5B, 64K for the 57B-A14B-Instruct, and up to 128K (131,072) tokens for the 7B-Instruct and 72B-Instruct. The flagship Qwen2-72B base model posts 84.2 on MMLU, 89.5 on GSM8K, 64.6 on HumanEval and 51.1 on MATH, beating contemporaries such as Llama 3 70B and Mixtral 8x22B on the Qwen team's reported numbers, while Qwen2-72B-Instruct reaches 86.0 on HumanEval and 91.1 on GSM8K.

On licensing, Qwen2 split its terms: the 0.5B, 1.5B, 7B and 57B-A14B models are released under Apache 2.0 for broad commercial use, while the 72B model keeps the more restrictive Tongyi Qianwen License. Qwen2 was the open-weight base that the much wider Qwen2.5 generation (September 2024) and later Qwen3 built on, and although it has been superseded, the weights remain freely downloadable from Hugging Face and ModelScope.

// Benchmarks

1. MMLU (5-shot, base 72B)84.2%
2. MMLU-Pro (base 72B)55.6%
3. GPQA (base 72B)37.9%
4. HumanEval (base 72B)64.6%
5. MBPP (base 72B)76.9%
6. GSM8K (base 72B)89.5%
7. MATH (base 72B)51.1%
8. HumanEval (72B-Instruct)86%
9. GSM8K (72B-Instruct)91.1%
10. MATH (72B-Instruct)59.7%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Permissive Apache 2.0 license on all sizes except the 72B, enabling broad commercial and self-hosted use
• Wide size range — from a 0.5B edge model to a 72B flagship and a 57B-A14B MoE — covering on-device through server deployments
• GQA across every size for faster inference and lower KV-cache memory
• Up to 128K-token context on the 7B-Instruct and 72B-Instruct variants
• Strong multilingual coverage: English, Chinese and 27 additional languages
• Competitive 2024 benchmark scores: 84.2 MMLU and 89.5 GSM8K on the 72B base model

// Best for

• Self-hosted, license-friendly LLM deployment via Apache 2.0 weights (sub-72B sizes)
• On-device and edge inference with the 0.5B and 1.5B models
• Long-context document processing using the 128K-context Instruct variants
• Multilingual chat and assistant applications across 29 languages
• Cost-efficient MoE serving with Qwen2-57B-A14B (14B activated parameters)
• Coding and math tasks via Qwen2-Instruct and the companion Qwen2-Math models

// How to access

// Qwen (open-weight) — every version

The full lineage of the Qwen (open-weight) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.