Last updated: 2026-07-01
Qwen is Alibaba Cloud's open-weight large language model family, spanning dense and Mixture-of-Experts architectures from 0.5 billion to over 1 trillion parameters. The flagship Qwen3-235B activates just 22 billion parameters per generation step, matching frontier coding and math benchmarks at a fraction of the compute cost, and most weights are free to self-host commercially.
About Qwen
Qwen (short for Tongyi Qianwen) is Alibaba Cloud's family of large language models and AI chat assistant, launched in beta in April 2023 and opened to the public that September. It reached 203 million monthly active users by February 2026, a 554% jump in a single month, and became the most-downloaded open-weight model family on Hugging Face with over 700 million downloads by January 2026, passing Meta's Llama in cumulative downloads. The model family spans 0.5 billion to over 1 trillion parameters across dense and Mixture-of-Experts architectures; the flagship Qwen3-235B activates only 22 billion parameters per step to keep inference costs down. Qwen Chat, the free consumer app at chat.qwen.ai, ships native clients for iOS, Android, Windows, and macOS. Developers reach the same models through Alibaba Cloud's Model Studio (DashScope), an OpenAI-compatible API, and most weights are released under an open license for royalty-free self-hosting. Alibaba shipped the latest release, Qwen 3.6-Plus, on April 2, 2026, continuing a rapid cadence that has kept Qwen competitive against Western frontier labs despite US chip export restrictions.
Screenshots
Pricing
Free tier: 1M tokens per model for 90 days after activating Model Studio. API pricing: Qwen-Flash $0.10/M input, $0.40/M output. Qwen-Plus $0.40/M input, $1.20/M output (non-thinking). Qwen-Max $1.20/M input, $6.00/M output (0-32K). 50% batch discount available. Qwen Chat consumer app is free to use.
Feature Comparison by Tier
| Feature | Qwen-Max API | Qwen-Plus API | Qwen-Flash API | Qwen Chat (Free) |
|---|---|---|---|---|
| Context window | 252K | 1M | 1M | 128K |
| Input token cost | $1.20/M | $0.40/M | $0.10/M | Free |
| Output token cost | $6.00/M | $1.20/M | $0.40/M | Free |
Key Features
- Hybrid Thinking Mode: Users can switch between fast non-thinking mode for quick answers and a slow deliberate reasoning mode for complex tasks, giving direct control over the speed-vs-depth tradeoff in a single toggle.
- 1M Token Context Window: Qwen 3.6 Plus Preview supports a 1-million-token context window with up to 65,536 output tokens, enabling full-codebase analysis or book-length document processing in a single call.
- Multimodal Input: Text, Image, Audio, Video: Qwen Chat handles text, image understanding via Qwen2.5-VL, audio input via Qwen2.5-Audio (speech, music, natural sounds), and video understanding, all within the same chat interface.
- 119+ Language Support: Qwen3 models support 119 languages and dialects, with Qwen3.5 extending coverage to 201 languages, making it one of the broadest multilingual LLM families available.
- Coding Across 92 Programming Languages: Qwen2.5-Coder, trained on 5.5 trillion tokens, supports 92 programming languages and achieved 69.6% on SWE-bench Verified, placing it among the top open-weight coding models.
- Qwen 3.6 MoE Architecture with 2x Throughput: Qwen 3.6 uses mature Mixture-of-Experts sparse activation with AITemplate kernel fusion, delivering approximately 2x inference throughput compared to Qwen 2.5 on comparable hardware.
- Qwen3.8-Max Preview: 2.4 Trillion Parameters: Alibaba previewed Qwen3.8-Max on July 19, 2026, a 2.4-trillion-parameter multimodal model the company positions as second only to Claude Fable 5 among frontier models, available via Token Plan, Qoder, and QoderWork at 10% of standard pricing ahead of a promised open-weight release.
- Qwen 3.6: 6 Size Tiers, Royalty-Free License: Qwen 3.6 was released in May 2026 across six model size tiers under an open license, enabling free commercial deployment for businesses matching compute budgets to task complexity.
Pros
- Most-downloaded open-weight model family globally with 700M Hugging Face downloads by January 2026, giving it a large community, broad fine-tune availability, and active third-party support.
- API pricing is significantly cheaper than frontier alternatives, roughly 40x less than GPT-4o on input tokens, making it practical for high-volume production workloads.
- Open license means no vendor lock-in for self-hosted deployments, unlike Llama's custom license which caps free commercial use at a fixed monthly-active-user threshold.
- Hybrid thinking mode lets developers choose reasoning depth per query, which is a concrete cost-saving lever not available in GPT-4o or base Claude models.
- The Plus Preview flagship ties GPT-5 mini on SWE-bench Verified at 72.4, showing frontier coding performance without frontier pricing.
Cons
- Qwen models refuse to answer questions about topics that conflict with the Chinese government's political positions, including Taiwan's government status and certain historical events, which limits use in journalism or political research.
- Image generation quality lags behind dedicated image models like Stable Diffusion or DALL-E 3, based on independent reviewer tests in 2025.
- Debugging and modifying existing codebases is weaker than writing new code from scratch, with reviewers noting fumbled refactors that a tool like Claude Sonnet handles better.
- The best API features and free-tier quotas are primarily available in the Singapore region; Global deployment has no free quota and Chinese Mainland endpoints are restricted to China-registered accounts.
Product Information
- Cloud
- Yes
- Self-Hosted
- Yes
- On-Premise
- Yes
- Languages
- English, Simplified Chinese, Traditional Chinese, Japanese, Korean, And 114+ more
- Training
- Official API Documentation, YouTube tutorials, Community Discord, GitHub examples, Hugging Face Hub guides
Frequently Asked Questions
What does Qwen actually cost?
Qwen Chat, the consumer chatbot, costs nothing. API access through Alibaba Cloud Model Studio bills per token: Qwen-Flash at $0.10 in and $0.40 out per million, Qwen-Plus at $0.40 in and $1.20 out, and Qwen-Max from $1.20 in and $6.00 out on its 0-32K tier. Batch jobs that are not time-sensitive get a 50% discount.
What do you get on Qwen's free tier?
Qwen Chat at chat.qwen.ai in full: text chat, image and video understanding, image generation, web search, and code execution, with no subscription. New Alibaba Cloud Model Studio accounts also get 1 million free API tokens per model for 90 days after activation. Beyond that allowance, API calls switch to standard per-token billing.
What should you use instead of Qwen?
Claude suits teams that want the strongest English reasoning and enterprise support. GPT-4o carries the broadest existing tool ecosystem for teams already built on it. DeepSeek matches or beats Qwen on raw coding benchmarks, which fits when Qwen's language and modality breadth is not needed.
Qwen or DeepSeek: which should you pick?
DeepSeek V4 leads on raw benchmarks, roughly 83.7% on SWE-bench and 99.4% on AIME, and undercuts Qwen-Max on output-token pricing. Qwen's edge is breadth: dozens of supported languages, multimodal input across text, image, audio and video, and an OpenAI-compatible Model Studio API that DeepSeek doesn't match. Teams chasing the top coding leaderboard score should pick DeepSeek; teams building multilingual or multimodal products should pick Qwen.
How do you set up Qwen?
For casual use, open chat.qwen.ai or install the iOS, Android, Windows, or macOS app and start chatting at no cost. For development, create an Alibaba Cloud account, activate Model Studio, and grab an API key from the DashScope console; the OpenAI-compatible endpoint means most existing SDKs work with just a base-URL change. Start on Qwen-Flash for testing before moving to Qwen-Plus or Qwen-Max in production.
Top Alternatives
- Claude: Claude goes deeper on English reasoning and enterprise support, while Qwen covers more languages at a fraction of the API cost.
- GPT-4o: GPT-4o has the broadest existing ecosystem; Qwen offers open-license self-hosting and far lower per-token costs.
- DeepSeek: DeepSeek leads on SWE-bench scores, while Qwen handles image, audio, and video input and broader language support in one tool.
HokAI guides covering Qwen
- What Qwen Is Actually For, Now It's Priced Below Claude and GPT-5.6: Alibaba priced Qwen3.8-Max at $2/$6 per million tokens, beating Claude and GPT-5.6, but its benchmarks are self-reported and the open-weights license isn't out.
- Best AI Chatbots in 2026: Pick by the Job, Not the Leaderboard: ChatGPT, Claude, Gemini, Mistral's Vibe, Poe and Character.ai compared on price and actual job fit, not raw model rank, with a decision table and 2026 pricing.