We build multilingual assistants, coding tools and document AI on Alibaba's Qwen models — open weights you can run anywhere, from a phone to your own data centre, or use through Alibaba Cloud.
Greetings rotate through some of the supported languages.
Qwen's open models use the same permissive licence as Kubernetes and Android. Here is how it compares with other popular open-weight licences.
| What you can do | Qwen (Apache 2.0) | DeepSeek (MIT) | Llama (community licence) |
|---|---|---|---|
| Commercial use | ✓ | ✓ | ~ |
| Modify and redistribute | ✓ | ✓ | ✓ |
| No user-count limit | ✓ | ✓ | ✕ |
| No “Built with…” branding requirement | ✓ | ✓ | ✕ |
| OSI-approved open-source licence | ✓ | ✓ | ✕ |
✓ yes · ~ with conditions · ✕ no. Apache 2.0 asks you to keep the licence and notices. A summary, not legal advice — check each model's licence file.
Qwen covers more sizes than almost any other model family. Scroll through the ladder to see where each size fits.
The smallest Qwen3.5 models run offline on phones and inside web apps. Good for autocomplete, simple classification, on-device privacy and places with poor connectivity.
Around 4B–14B parameters: fast, cheap and surprisingly capable for chat, extraction, summaries and tool calls, especially after fine-tuning.
Qwen3 32B and the efficient Qwen3.5 35B-A3B (only about 3B active per token) deliver strong quality on one serious GPU.
Qwen3 235B-A22B uses 22B active parameters per token: near-frontier quality on a multi-GPU server you control.
The open flagship released in February 2026: 397B total, about 17B active, 262K-token native context and support for 201 languages. Also available through Alibaba Cloud's API.
Beyond the general chat models, Qwen has specialists that beat general models on their own jobs.
Code generation, repair and agentic coding across large repositories.
Qwen3-Coder 480B-A35B and smaller sizesReads screenshots, charts, invoices, forms and photos, and answers questions about them.
Image and document understandingListens, watches and talks: voice assistants and video understanding in one model.
Speech in, speech outTurns text into vectors for multilingual search, RAG and recommendations.
Search and retrievalQwen3.5 supports 201 languages and dialects — a big advantage for products in India, Asia, the Middle East and Africa.
Qwen3.5 is trained to read and write a very wide range of languages, not just English and Chinese.
Serve customers in Hindi, Arabic, Spanish and English from the same deployment.
Answer users directly in their language instead of translating twice.
Qwen can answer instantly or reason first, and you can cap the thinking budget. Slide to see the trade-off on a real business question.
“Plan A or Plan B for our 24-month software contract?”
Illustrative example.
We set the budget per feature: off for quick replies, high for calculations, planning and code.
Pick the route that fits your data rules, volume and team. We can start on an API and move to self-hosting later without rewriting your app.
Managed API access to Qwen models, including the larger Plus and Max models that are API-only.
Apache 2.0 weights on your own GPUs or private cloud — your data never leaves your control.
Third-party hosts and routers serve open Qwen models per token in various regions.
Where Qwen's languages, sizes and licence give a real advantage.
We are model-neutral and often combine several behind one API.
| Model family | Strengths | When we choose it |
|---|---|---|
| Qwen | Apache 2.0 open weights, every size, broad language coverage, coding and vision variants | Multilingual products, permissive licensing, any-size deployment |
| Meta Llama | Open weights with wide tool support (community licence) | Private deployments, fine-tuning |
| DeepSeek | Low-cost open reasoning (MIT) | Cost-sensitive reasoning |
| OpenAI | Hosted voice, agents and images | Multimodal products, fastest launch |
| Anthropic Claude | Long documents, coding, dependable agents | Contracts, research, coding agents |
| Google Gemini | Multimodal with Google grounding | Video, voice, Google Cloud apps |
Straight answers to what clients ask us most.
Qwen is Alibaba's family of large language models. Many are released as open weights under the Apache 2.0 licence, from tiny models that run on a phone to the 397B-parameter Qwen3.5 flagship, plus coding, vision, speech and embedding variants. Larger Plus and Max models are available through Alibaba Cloud's API.
The open Qwen3 and Qwen3.5 models are released under Apache 2.0, one of the most permissive licences: commercial use, modification and redistribution are allowed with attribution. Always check the licence file of the exact model you use.
Released in February 2026, Qwen3.5 brought a 397B-A17B open flagship with a hybrid linear-attention and mixture-of-experts design, 262K-token native context and support for 201 languages, plus small 0.8B and 9B models and an efficient 35B-A3B.
Language coverage is one of Qwen's strengths. Quality still varies by language and task, so we test on real examples in your languages, and fine-tune or add retrieval where needed.
Yes. The open weights run with vLLM, SGLang, Ollama or llama.cpp on your GPUs or private cloud, and small models even run on laptops and phones. We size the hardware and set up monitoring.
Qwen3 introduced hybrid thinking: the same model can answer instantly or reason step by step first, and you can limit how much it thinks. We tune this per feature to balance accuracy, speed and cost.
Both are excellent open-weight families. Qwen's Apache 2.0 licence is more permissive than Llama's community licence, and its language coverage is broad; Llama has the widest ecosystem in the West. We often test both on your task.
Tell us what you want to build, which languages you need and where it must run. We reply within 4 business hours with honest advice.