Model Details
Qwen2 72B Instruct
Quick answer: Qwen2 72B Instruct is Alibaba's June 2024 flagship open-weight model, scoring 89.5% on GSM8K, 82.3% on MMLU, 64.4% on MMLU-Pro, and 86.0% on HumanEval. With a 128K context window and Qianwen License, it was among the strongest open-weight 72B models at launch and has since been succeeded by Qwen2.5 and Qwen3.
At a Glance
Where Qwen2 72B leads
- 89.5% GSM8K — competitive math reasoning at 72B open-weight scale
- 86.0% HumanEval — strong code completion
- 82.3% MMLU — competitive general knowledge
- 128K context window
- Strong Chinese-language capability (designed for bilingual EN/ZH)
- Open weights on Hugging Face
Where it lags
- Qianwen License (not Apache 2.0) — commercial restrictions apply
- Superseded by Qwen2.5 72B and Qwen3 32B/235B
- 64.4% MMLU-Pro — below newer frontier open models
- Knowledge cutoff March 2024
Best for: Existing Qwen2 72B deployments; strong Chinese-English bilingual tasks; foundation for Qwen2-based fine-tunes.
What Qwen2 72B Instruct Is
Qwen2 72B Instruct (released June 7, 2024) was Alibaba's state-of-the-art open-weight model at launch, competing directly with Llama 3 70B and GPT-4o-mini on standard benchmarks. It demonstrated particularly strong performance in mathematics (89.5% GSM8K) and code (86.0% HumanEval) relative to its parameter count.
The Qwen2 family was designed with bilingual Chinese-English capability as a first-class requirement — outperforming Llama 3 on Chinese benchmarks while remaining competitive on English tasks. This made it a leading choice for multilingual applications in 2024.
For new deployments in 2025, Qwen3 32B offers better performance across all benchmarks under Apache 2.0 with hybrid thinking capability. Qwen2 72B Instruct is maintained for backwards compatibility and existing integrations.
Specifications
| Field | Value |
|---|---|
| Organization | Alibaba / Qwen |
| Parameters | 72B (dense) |
| Context window | 128,000 tokens |
| License | Qianwen License |
| HuggingFace | Qwen/Qwen2-72B-Instruct |
| Release date | June 7, 2024 |
| Knowledge cutoff | March 2024 |
| Modality | Text only |
Pricing
Qwen2 72B Instruct is available as open weights under the Qianwen License on Hugging Face. Available via Alibaba Cloud DashScope and third-party providers.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| GSM8K | 89.5% | Benchgen evaluation | 2025-07 |
| MMLU | 82.3% | Benchgen evaluation | 2025-07 |
| MMLU-Pro | 64.4% | Benchgen evaluation | 2025-07 |
| HumanEval | 86.0% | Benchgen evaluation | 2025-07 |
| MATH | 59.7% | Benchgen evaluation | 2025-07 |
| BigCodeBench | 20.6% | Benchgen evaluation | 2025-07 |
Qwen2 72B vs Alternatives
| Model | MMLU | MMLU-Pro | HumanEval | License |
|---|---|---|---|---|
| Qwen2 72B Instruct | 82.3% | 64.4% | 86.0% | Qianwen |
| Llama 3.1 70B Instruct | 86.0% | — | — | Llama 3.1 |
| Qwen3 32B | — | — | — | Apache 2.0 |
| DeepSeek-V3 | 88.5% | — | — | MIT |
Qwen3 32B (April 2025, Apache 2.0) is preferred for new Qwen deployments: better performance at lower parameter count, hybrid thinking, fully permissive license. Qwen2 72B is appropriate for existing workflows.
Related Models & Benchmarks
- Compare: Qwen3 32B · Llama 3.1 70B Instruct · DeepSeek-V3
- Benchmarks: MMLU · MMLU-Pro · GSM8K
Frequently Asked Questions
What is Qwen2 72B Instruct?
Qwen2 72B Instruct is Alibaba's June 2024 open-weight 72B model, scoring 89.5% GSM8K, 82.3% MMLU, and 86.0% HumanEval with 128K context under the Qianwen License.Is Qwen2 72B open source?
The weights are open under the Qianwen License. Commercial use requires agreeing to Alibaba's license terms — it is not Apache 2.0 or MIT. For fully permissive open weights, consider Qwen3 32B (Apache 2.0).Should I use Qwen2 72B or Qwen3?
Qwen3 32B (Apache 2.0, April 2025) for new deployments: better benchmarks, hybrid thinking, fully permissive license. Qwen2 72B Instruct is appropriate for existing integrations.Specs from Alibaba's official Qwen2 release (June 2024) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.