Qwen 2.5-72B Instruct
Open weightsby Alibaba Cloud·China·Released
Open-weights 72B from the Qwen 2.5 family — Apache 2.0, runs on a single 8×A100 node.
About this model
Qwen 2.5-72B Instruct (September 2024) is the largest open-weights model in the Qwen 2.5 family — released under Apache 2.0, runs on a single 8×A100 or 8×H100 node, and has become the default open-weights choice for teams that find Llama 3.3 70B's Community License too restrictive.
On benchmarks it's broadly comparable to Llama 3.3 70B, with stronger Chinese-language performance and competitive English scores. The Apache 2.0 license is genuinely permissive — no MAU caps, no commercial-use restrictions.
Strengths
- •Apache 2.0 — most permissive license among frontier-adjacent models
- •No MAU restrictions (unlike Llama Community License)
- •Strong Chinese + English bilingual capability
- •Mature fine-tuning ecosystem (LoRA / QLoRA support is excellent)
Limitations
- •Closed Qwen 2.5-Max outperforms it on hardest tasks
- •Smaller activated parameter count limits hardest reasoning
- •Less benchmark coverage on English-only tasks than Llama family
When to use it
- →On-prem deployments under strict licensing requirements
- →Fine-tuning for company-specific corpora
- →Bilingual applications (Chinese-English)
- →Cost-sensitive serving where Apache 2.0 matters
Architecture & training
72B-parameter dense transformer. Pretrained on Alibaba's TaoBao-affiliated GPU clusters. The Qwen 2.5 technical report describes the training mix as ~36% Chinese, ~36% English, ~14% code, with the remainder being other languages and structured data. Post-training uses RLHF and Direct Preference Optimisation (DPO).
Benchmarks
| Benchmark | Score | Bar |
|---|---|---|
| MATH | 83.1 | |
| MMLU | 86.1 | |
| HumanEval | 86.6 |