Curiosity · AI Model
Qwen 2.5 72B Instruct
Qwen 2.5 72B Instruct is Alibaba Cloud's September 2024 open-weights flagship — a 72B dense transformer that matched Llama 3.1 405B on several benchmarks while being roughly one-sixth the size. Apache 2.0 licensed and with broad multilingual coverage, it defined the Qwen line's credibility in Western open-source benchmarking.
Model specs
- Vendor
- Alibaba
- Family
- Qwen 2.5
- Released
- 2024-09
- Context window
- 131,072 tokens
- Modalities
- text
- Input price
- $0.35/M tok
- Output price
- $0.4/M tok
- Pricing as of
- 2026-04-20
Strengths
- Apache 2.0 license — fully permissive commercial use
- Matches or beats Llama 3.1 70B on most public benchmarks
- 128K context with strong needle-in-haystack retrieval
- Best-in-class Chinese, Japanese, Korean language quality
Limitations
- Superseded by Qwen 3 dense and MoE variants in 2025
- Trails DeepSeek V3 and Llama 4 Maverick on top-end reasoning
- Enterprise adoption in US / EU slowed by geopolitical caution
- Tool-use and agentic fine-tuning weaker than closed frontier
Use cases
- Multilingual production assistants and chatbots
- Fine-tuning base across dozens of community derivatives
- RAG pipelines with 128K context
- Research baselines for open-weights mid-flagship comparisons
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MMLU | ≈85% | 2024-09 |
| HumanEval | ≈86% | 2024-09 |
| MATH | ≈83% | 2024-09 |
Frequently asked questions
Is Qwen 2.5 72B still worth using?
As a stable, well-supported open-weights flagship from 2024, yes — especially for Apache 2.0 deployments or multilingual workloads. For bleeding-edge quality, look at Qwen 3 or DeepSeek V3.
How does Qwen 2.5 72B compare to Llama 3.1 70B?
Qwen 2.5 72B typically beats Llama 3.1 70B on math and Chinese-language benchmarks, matches it on English reasoning, and ships with a more permissive Apache 2.0 license.
What sizes does the Qwen 2.5 family cover?
Qwen 2.5 comes in 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B dense variants, plus specialized Qwen 2.5 Coder and Qwen 2.5 Math models.
Sources
- Qwen — Qwen 2.5 announcement — accessed 2026-04-20
- Hugging Face — Qwen/Qwen2.5-72B-Instruct — accessed 2026-04-20