Curiosity · AI Model
Qwen 3
Qwen 3 is Alibaba Cloud's 2025 open-weights flagship family — a sweep of dense and Mixture-of-Experts models from 0.6B all the way to Qwen3-235B-A22B, all released under Apache 2.0. It unified Alibaba's chat, coding, and reasoning lines and is one of the most benchmarked open families of 2025.
Model specs
- Vendor
- Alibaba
- Family
- Qwen 3
- Released
- 2025-04
- Context window
- 128,000 tokens
- Modalities
- text
- Input price
- $0.2/M tok
- Output price
- $0.6/M tok
- Pricing as of
- 2026-04-20
Strengths
- Apache 2.0 across the family — no community-license caveats
- Covers the full size spectrum from edge to frontier-scale MoE
- Top-tier Chinese and multilingual performance
- Native thinking / non-thinking hybrid mode for reasoning control
Limitations
- Western-market enterprise adoption slowed by geopolitical concerns
- Community smaller than Llama in Western toolchains
- Largest Qwen3-235B still trails Claude Opus and GPT-5 on top-end agentic tasks
- Safety alignment less thoroughly audited than Western frontier releases
Use cases
- Multilingual production assistants, especially Chinese and SE Asian languages
- Research baselines where Apache 2.0 licensing matters
- Fine-tuning base across size ranges (0.6B to 235B)
- Agentic workflows needing the larger Qwen3-235B MoE
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MMLU-Pro | ≈79% | 2025-04 |
| HumanEval | ≈89% | 2025-04 |
| MATH-500 | ≈88% | 2025-04 |
Frequently asked questions
What is Qwen 3?
Qwen 3 is Alibaba Cloud's 2025 open-weights LLM family, covering dense 0.6B–32B sizes and MoE variants up to Qwen3-235B-A22B. All models ship under Apache 2.0.
What makes Qwen 3 special?
Three things: full Apache 2.0 license across the family, best-in-class Chinese and multilingual coverage, and a hybrid thinking/non-thinking mode that lets a single model switch between fast direct answers and long reasoning traces.
Is Qwen 3 safe for enterprise use?
The license permits it, but some Western enterprises avoid Chinese-origin models for data-sovereignty and supply-chain reasons. Evaluate based on your risk tolerance — the weights themselves are public and auditable.
Sources
- Qwen — Qwen 3 announcement — accessed 2026-04-20
- Hugging Face — Qwen/Qwen3 collection — accessed 2026-04-20