Qwen
The Chinese open-weight family, among the most complete and versatile.
Olya's profile
- What sets it apart
- Qwen is a sprawling family — over 145 models under a single API — and, crucially, open-weight under the Apache 2.0 license, from tiny models up to the large 235-billion-parameter MoE: you can download them and run them on your own hardware. It covers 119 languages, is multimodal (text, images, audio, video) and has a reasoning mode you can switch on. The trade-off is that the newest top-tier 'Max' models are now closed and API-only.
- Strengths
- It's cheap and it delivers: on Alibaba's cloud the MoE models run around $0.40/$2.40 per million tokens, a fraction of Western flagships, and the open versions remove vendor lock-in entirely. Its language coverage and specialized variants — code, vision, audio — are among the most complete out there, and the chat app stays free with no sign-up.
- When to use it
- It pays off when you want to self-host for privacy or cost, when you need Asian languages or broad multilingual support, or when you're handling high request volumes without blowing up the bill. It's also a solid base for anyone wanting to experiment with fine-tuning without paying license fees.
- When to avoid it
- If you want the absolute best on the hardest reasoning tasks, the flagships from OpenAI, Google and Anthropic are still ahead — and Qwen's own 'Max' models are, ironically, just as closed as theirs. The ecosystem and docs revolve around Alibaba's Chinese cloud, which adds bureaucratic friction and data-compliance doubts for anyone operating in the EU. The free developer tier was shut down in April 2026.