The things people actually ask about Tongyi Qwen licensing, cost, which model to pick, and how it compares plus a glossary for the jargon and the sources behind this site.
They are the same thing. 通义千问 (Tongyi Qianwen) is the Chinese product name Alibaba Cloud announced in April 2023; Qwen is the international name used for the models, the code repositories and the API. Tongyi Lab is the research team inside Alibaba Cloud that builds them. You will see all three used interchangeably.
Partly. Most Qwen models are released as open weights - you can download and run them at no licence cost, which is genuinely free apart from your own compute. The consumer Qwen Chat and Qwen App are free to use. The hosted Max, Plus and Flash tiers are paid per token, though Alibaba Cloud offers around a million free tokens to new Model Studio users.
Usually yes, but the answer is per model, not per family. Apache 2.0 releases - which cover most of the open lineup including Qwen3.8-27B - permit commercial use, modification and redistribution with attribution. Some models carry the source-available Tongyi Qianwen licence with conditions, some research checkpoints are explicitly non-commercial, and Qwen3.8-2.4T-A95B ships under its own bespoke licence. Read the LICENSE file in the specific repository, and have counsel review it if the deployment matters.
It is the active parameter count in a mixture-of-experts model. Qwen3.8-2.4T-A95B holds 2.4 trillion parameters in total but routes each token through only about 95 billion of them. Memory requirements track the total; compute cost and speed track the active figure. That is why a 2.4T model can be served at all - but it still has to fit in memory, so it remains a cluster-scale model.
A mode where the model generates internal reasoning before its final answer, which improves accuracy on maths, code and multi-step problems at the cost of latency and tokens. Qwen3 introduced it as a toggle; by the 3.8 generation, Qwen3.8-27B has it on by default and switchable per query, while Qwen3.8-2.4T-A95B always thinks and cannot be switched off. Hosted Max models expose a reasoning_effort parameter (low, medium, xhigh) instead of a binary switch.
On published August 2026 figures, Qwen3.8-Max leads on agentic computer use (86.1 OSWorld-Verified) and instruction following (82.8 IFBench), sits within a couple of points of the leaders on terminal tasks and graduate-level science, and trails Claude Fable 5 by roughly twelve points on SWE-bench Pro. Its bigger structural advantage is different: no other frontier-class family publishes weights at this scale, so you can run it in your own environment. See the benchmarks page for the full comparison and the caveats that go with it.
That depends entirely on how you deploy it. Using the hosted API sends your data to Alibaba Cloud, governed by their service terms and the region you selected - which is a meaningful consideration for regulated data and in some jurisdictions a policy question rather than a technical one. Self-hosting the open weights keeps data on infrastructure you control, which is a large part of why open-weight models are adopted in healthcare, finance and government. Treat model outputs themselves as untrusted input in any agentic pipeline, regardless of vendor.
Very often. In 2026 alone the numbered generations went 3.5 (February) → 3.6 (April) → 3.7 (May–June) → 3.8 (July–August), with separate image and audio releases in between. Pin exact model IDs in production, and re-benchmark quarterly rather than assuming your choice from six months ago still holds - small dense models in particular have improved fast enough to displace much larger ones.
The signals cut both ways, and reasonable people read them differently. Against: 2026 brought proprietary Plus and Max tiers, a generation (3.7) with no open release at all, and the departure of Qwen division head Lin Junyang in March. For: Alibaba has published over a hundred open checkpoints, kept shipping Apache 2.0 models throughout, and in August 2026 released Max-class weights for the first time - something no competitor at that tier has done. The honest summary is a deliberate two-track strategy rather than a retreat, with the boundary between tracks moving release by release.
Yes, for the open-weight models. Downloaded checkpoints run entirely locally through vLLM, SGLang, llama.cpp, Ollama or MLX with no network connection. The hosted Max, Plus and Flash tiers cannot - they exist only as an API. If offline operation is a requirement, it constrains you to the open line, which is one of the main reasons that line exists.
The Qwen3 generation was trained across 119 languages and dialects, with unusually strong Chinese and English. Qwen3.5-Omni-Plus handles 113 languages for speech, Qwen-Audio-3.0-TTS covers 16 for generation, and Qwen-Image-3.0 renders readable text inside images in 12. Quality varies considerably by language - test on yours rather than trusting the headline count.
| Term | Meaning |
|---|---|
| MoE (mixture of experts) | An architecture that routes each token through a small subset of specialised sub-networks, so a very large model can be run at the cost of a much smaller one. |
| Active parameters | The parameters actually used per token in an MoE model - the figure after the “A” in names like 2.4T-A95B. Drives speed and compute cost. |
| Context window | How much text the model can consider at once, measured in tokens. Qwen's current flagships reach roughly one million. |
| YaRN | A technique for extending a model's usable context beyond what it was trained on. Qwen3.8-27B uses it to go from 262K to 1M. |
| Gated DeltaNet | A linear-attention layer type. Qwen's recent models mix roughly 75% linear-attention layers with 25% full attention to make long context affordable. |
| Open weights | The trained parameters are downloadable. Not the same as open source - training data and code are generally not published, and licence terms vary. |
| Instruct vs Base | Base is the raw pretrained model; Instruct has been post-trained to follow instructions and is what you almost always want. |
| Agentic | Using tools, running commands and taking multi-step actions toward a goal rather than producing a single answer. The organising theme of Qwen's 2026 releases. |
| Model Studio / DashScope | Alibaba Cloud's platform and API for hosted Qwen models, with separate international and mainland-China deployments. |
| ModelScope | Alibaba's model hub - the Chinese-ecosystem counterpart to Hugging Face, where Qwen weights are published in parallel. |
Compiled from official Alibaba announcements and model cards plus independent reporting, as of 18 August 2026. Where sources disagree on an exact release date announcement versus weight publication, for instance this site notes the range rather than picking one.