Last updated: 2026-07-01
Qwen is Alibaba's open-source LLM series, available free under Apache 2.0 for commercial use. Alibaba previewed its new flagship Qwen3.8-Max on July 19, 2026, a 2.4-trillion-parameter multimodal model positioned as second only to Claude Fable 5. Qwen 3.6 offers a 1M token context window, 119+ language support, and six model size tiers.
About Qwen
Qwen (short for Tongyi Qianwen) is Alibaba Cloud's family of large language models and chat assistant, first launched in beta in April 2023 and opened to the public in September 2023. It reached 203 million monthly active users by February 2026, a 554% spike in a single month, and became the most-downloaded open-weight model family on Hugging Face with over 700 million downloads by January 2026, surpassing Meta's Llama in cumulative downloads. The model family spans 0.5 billion to over 1 trillion parameters, with both dense and mixture-of-experts (MoE) architectures. The flagship Qwen3-235B model activates only 22 billion parameters per generation step, keeping inference costs low while delivering competitive results. A standout design choice is the hybrid thinking mode: users can toggle between fast non-thinking mode for quick answers and a slower deliberate reasoning mode for complex math, code, or analysis tasks. The Qwen 3.6 Plus Preview extends context to 1 million tokens and matches GPT-5 mini on SWE-bench Verified at 72.4. Qwen Chat is the consumer-facing product, available on web at chat.qwen.ai, and via native apps for iOS, Android, Windows, and macOS. It handles text chat, document processing, image understanding, image generation, video understanding, web search, and code execution in a single interface. The underlying Qwen2.5-Coder model was trained on 5.5 trillion tokens and supports 92 programming languages. API access runs through Alibaba Cloud's Model Studio (DashScope), which also offers an OpenAI-compatible endpoint. Qwen-Flash costs $0.10 per million input tokens, Qwen-Plus costs $0.40 per million input tokens, and Qwen-Max starts at $1.20 per million input tokens. All new API accounts get 1 million free tokens per model valid for 90 days. Over 90,000 enterprises have adopted Qwen models via Model Studio. Qwen models are released under Apache 2.0, letting developers self-host or fine-tune without licensing restrictions. The Qwen Agent framework provides tooling for building multi-step AI workflows. Alibaba released Qwen 3.6-Plus on April 2, 2026, adding stronger coding and agent capabilities, continuing a rapid release cadence that has kept Qwen competitive against Western frontier models 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, Apache 2.0: Qwen 3.6 was released in May 2026 across six model size tiers under Apache 2.0, enabling royalty-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: Qwen-Flash at $0.10/M input tokens is roughly 40x cheaper than GPT-4o, making it practical for high-volume production workloads.
- Apache 2.0 license means no vendor lock-in for self-hosted deployments, unlike Llama's custom license that restricts commercial use above 700M monthly active users.
- 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.
- Qwen 3.6 Plus Preview 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 is Qwen and who built it?
Qwen is Alibaba Cloud's family of large language models and AI chat assistant, first launched in beta in April 2023 and opened to the public in September 2023 under the Chinese name Tongyi Qianwen. Alibaba Cloud, the cloud computing arm of Alibaba Group, develops and operates Qwen through its Tongyi Lab research unit. By February 2026 Qwen Chat reached 203 million monthly active users, a 554% jump in a single month. The model family spans 0.5 billion to over 1 trillion parameters across dense and Mixture-of-Experts architectures, with the flagship Qwen3-235B activating only 22 billion parameters per step. Qwen models became the most-downloaded open-weight family on Hugging Face, passing 700 million downloads by January 2026 and overtaking Meta's Llama in cumulative downloads. Most model weights are released under Apache 2.0, so companies can self-host or fine-tune without licensing fees.
How much does Qwen cost in 2026?
Qwen Chat, the consumer chatbot at chat.qwen.ai, is free to use with no subscription required. API access runs through Alibaba Cloud Model Studio (DashScope) and is billed per token: Qwen-Flash costs $0.10 per million input tokens and $0.40 per million output tokens, Qwen-Plus costs $0.40 per million input tokens and $1.20 per million output tokens, and Qwen-Max starts at $1.20 per million input tokens and $6.00 per million output tokens for the 0-32K context tier. Cached input tokens are billed at roughly half the standard input rate across all three models. New Model Studio accounts receive 1 million free tokens per model, valid for 90 days after activation. A 50% batch-processing discount is also available for non-time-sensitive jobs. Self-hosting the open-weight models under Apache 2.0 avoids per-token fees entirely but requires your own GPU infrastructure.
What does Qwen do that competitors don't?
Qwen's hybrid thinking mode lets users toggle between a fast non-thinking mode for quick answers and a slower deliberate reasoning mode for complex math, code, or analysis, switchable per request rather than locked to a separate model. The Qwen 3.6 Plus Preview extends context to 1 million tokens and matches GPT-5 mini on SWE-bench Verified at 72.4, while the flagship Qwen3-235B Mixture-of-Experts model activates only 22 billion of its 235 billion parameters per generation step, keeping inference costs low. Qwen2.5-Coder, trained on 5.5 trillion tokens, supports 92 programming languages. Almost the entire model lineup, from 0.5B to over 1 trillion parameters, ships under Apache 2.0, letting enterprises self-host for data sovereignty without licensing restrictions, a combination of scale, openness, and per-token pricing that few Western labs match in 2026.
How does Qwen compare to DeepSeek?
Both Qwen and DeepSeek ship open-weight models under permissive licenses (Apache 2.0 for Qwen, MIT for DeepSeek) and both are production-ready for enterprise use in 2026. On raw benchmarks, DeepSeek V4 leads with around 83.7% on SWE-bench and 99.4% on AIME, but Qwen 3.6-35B-A3B leads the sub-40B weight class with 86.0% on GPQA and 92.7% on AIME 2026, making it a strong single-GPU option. DeepSeek V4-Pro is cheaper on coding-heavy output workloads at roughly $3.48 per million output tokens versus Qwen-Max's $6.00. Qwen's edge is breadth: 119+ languages, a 1M-token context option, multimodal input covering text, image, audio and video, and the OpenAI-compatible Model Studio API. Teams choosing between them often pick DeepSeek for raw coding benchmark scores and Qwen for multilingual, multimodal, and consumer-app coverage.
Is Qwen free to use?
Yes. Qwen Chat at chat.qwen.ai is completely free, with no subscription tier, covering text chat, document processing, image understanding, image generation, video understanding, web search, and code execution in one interface, plus native apps for iOS, Android, Windows, and macOS. For developers, Alibaba Cloud Model Studio gives every new account 1 million free tokens per model for 90 days after activation. Beyond that allowance, API calls are billed per token, starting at $0.10 per million input tokens for Qwen-Flash. The underlying model weights, from 0.5 billion to over 1 trillion parameters, are released under Apache 2.0, so organizations can download and self-host them at zero licensing cost, paying only for their own compute. Over 90,000 enterprises have adopted Qwen models through Model Studio as of 2026.
Who is Qwen best for and who should avoid it?
Qwen is best for developers and enterprises that want a free, capable chat assistant plus an open-weight model family they can self-host under Apache 2.0, particularly teams in regions where US-based models face cost, latency, or access restrictions. Its 119+ language support and 92-language code coverage suit multilingual products and global support teams, and its tiered model sizes (0.5B to 235B+ active MoE) let teams match a model to their hardware budget. Qwen may not suit organizations that require data residency strictly outside Chinese-affiliated cloud infrastructure for the hosted API, or teams that need the absolute top score on Western coding leaderboards, where DeepSeek V4 and GPT-5.5-class models currently score higher on SWE-bench. Teams with strict US-government compliance requirements should evaluate Alibaba Cloud's compliance posture before deploying the hosted API in production.
Does Qwen work for coding and agentic tasks?
Yes. Qwen2.5-Coder was trained on 5.5 trillion tokens and supports 92 programming languages, and the Qwen3-235B flagship matches GPT-5 mini on SWE-bench Verified at 72.4 in its Qwen 3.6 Plus Preview configuration with a 1 million token context window. The Qwen Agent framework provides tooling for building multi-step AI workflows, and the qwen-code CLI tool lets developers run Qwen models in agentic coding loops similar to Claude Code or Cursor's agent mode. The hybrid thinking mode is useful here: non-thinking mode handles quick autocomplete-style edits, while thinking mode is better for multi-file refactors or debugging. Some GitHub issues on QwenLM/qwen-code report tool-calling errors when an assistant message with tool_calls isn't immediately followed by matching tool responses; pinning to a stable qwen-code release version avoids most of these.
Does Qwen train on user data and is it safe for business use?
Alibaba Cloud states that Model Studio API usage is governed by its own data-handling terms separate from the consumer Qwen Chat app, and Alibaba Cloud holds ISO 27001 and SOC 2 compliance certifications for its cloud infrastructure as of 2025. For the open-weight models released under Apache 2.0, from 0.5B to over 1 trillion parameters, organizations can download and run them entirely on their own infrastructure, meaning no data ever leaves their environment, which is the preferred path for regulated industries. Over 90,000 enterprises had adopted Qwen models via Model Studio as of early 2026. Businesses with strict data-sovereignty requirements should review Alibaba Cloud's specific data-processing terms for the hosted API and consider self-hosting the Apache 2.0 weights instead if cross-border data transfer is a concern.
Top Alternatives
- Claude: Pick Claude for maximum reasoning in English; pick Qwen if you need 119+ languages and lower API costs.
- GPT-4o: Pick GPT-4o for ecosystem lock-in; pick Qwen for Apache 2.0 self-hosting and 40x cheaper tokens.
- DeepSeek: Pick DeepSeek for raw coding benchamkrs; pick Qwen for broader multimodal and multilingual coverage.
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.