Qwen

Qwen is the family name for the large language and multimodal models developed by Alibaba, first announced in April 2023 as Tongyi Qianwen and later branded Qwen for international use. It spans small models that run on a phone, open-weight models in the tens and hundreds of billions of parameters, and proprietary systems that Alibaba sells through its cloud 1.

Qwen's importance lies in the combination of scale and openness. Alibaba has released the weights of most of its models under permissive licences, including the flagship 235-billion-parameter mixture-of-experts version of Qwen3, and those weights are distributed through ModelScope, the Chinese model hub, as well as Hugging Face. That made Qwen a common foundation for developers who want a capable model they can host themselves 2 3.

The family is also the most visible example of China's open-weight strategy. Where American labs have largely kept their best models behind APIs, Alibaba and several Chinese competitors have published weights, which changes both the economics of AI and the policy debate around it. This entry covers the models, their training, benchmarks, licensing, pricing and the criticism they have attracted.

Qwen is not a single model but a set of related ones. The core text models come in dense sizes from 0.5 billion to 72 billion parameters and in mixture-of-experts versions such as Qwen3-235B-A22B, which has 235 billion total parameters and activates 22 billion per token. Separate lines cover vision and language (Qwen-VL), audio (Qwen-Audio), full multimodality (Qwen-Omni) and code (Qwen-Coder) 1.

The naming has been inconsistent. Tongyi Qianwen is the Chinese product name; Qwen is the model name used in research and internationally; Qwen Chat is the consumer interface; and Model Studio is the cloud platform that sells API access. The result is that the same model can appear under several names, and older documentation is not always clear about which version it describes 4.

The open releases are the ones that matter most outside China. Alibaba publishes model cards, technical reports and evaluation scripts, and the weights can be downloaded without a commercial relationship. The largest models, including Qwen3-Max, are not released and are available only through the API, which lets the company keep a capability lead while the open models trail it by months 2.

Qwen is also embedded in Alibaba's own products: the Tongyi assistant, the DingTalk workplace app, the Amap maps service and parts of Taobao. That internal deployment gives the team usage data and revenue that pure research labs lack, and it means the models are optimised for consumer tasks in Chinese as well as for developer benchmarks 4.

Tongyi Qianwen was announced in April 2023, and the first open models followed in August, when Alibaba released Qwen-7B and a vision-language version. These early releases established the pattern: a proprietary flagship for cloud customers, open versions for the community, and technical reports describing the training 5.

Qwen2 arrived in June 2024 with sizes from 0.5 billion to 72 billion parameters, improved multilingual coverage and a permissive licence for most models. Qwen2.5 followed in September 2024, trained on 18 trillion tokens, with stronger mathematics and coding and a long-context variant. Qwen2.5-Max, a proprietary mixture-of-experts model, was released in January 2025 6.

Qwen3, released in April 2025, was the generation that established the family internationally. It introduced a hybrid mode in which a single model can answer directly or produce a reasoning trace, with the mode selectable at inference. The open release covered six dense sizes and two mixture-of-experts models, all under Apache 2.0, and the company reported competitive results against contemporary reasoning models 2.

Later releases extended the line. Qwen3-Coder, a 480-billion-parameter mixture-of-experts model for software engineering, arrived in mid-2025; Qwen3-Max, a proprietary model with more than a trillion parameters, followed in September; and Qwen3-Next and Qwen3-VL extended the architecture and multimodal range. A consumer app, Qwen, launched in November 2025 1.

The Qwen3 technical report gives more detail than most Chinese labs publish. The models are decoder-only transformers; the mixture-of-experts versions use 128 experts with eight activated per token, and the dense models range from 0.6 billion to 32 billion parameters. Pre-training used 36 trillion tokens covering 119 languages, with a curriculum that moves from general text to higher-quality and more specialised material 6.

The hybrid thinking design is the most distinctive feature. Rather than shipping separate reasoning and chat models, Qwen3 was trained so that one model can operate in either mode, controlled by a token or a parameter. The model was first trained on reasoning tasks and then adapted for general conversation, a sequence the team says preserves reasoning ability while restoring fluency 6.

Post-training included reinforcement learning on verifiable tasks and distillation from the large models into the small ones. Alibaba published the distillation recipes alongside the weights, which is unusual and useful: it lets other groups reproduce the small models rather than guessing at the process. The company has not disclosed its full data mixture, filtering rules or compute budget 2.

Safety training is described in model cards rather than in the technical report. Qwen models are trained to comply with Chinese content regulations, which require the removal of material deemed harmful or politically sensitive, and the cards describe red-teaming and refusal behaviour. Independent researchers have documented that the models refuse topics that Western models discuss, a difference that reflects the regulatory environment rather than an accident of training 5.

Qwen's strongest documented areas are multilingual text, mathematics, coding and Chinese-language tasks. The larger models handle translation between Chinese and English particularly well, and the coding models are competitive with specialised proprietary systems on repository-level work. The small dense models are unusually capable for their size, which is why they appear in on-device and edge deployments 2.

Multimodal versions add image and document understanding, and Qwen-Omni accepts audio and video as well. The vision models are used for OCR, chart reading and document extraction in Chinese business settings, where they have an advantage over models trained mostly on English data. Video understanding remains limited to short clips and sampled frames 1.

The limits are those of open models generally, plus some specific ones. The largest open versions need substantial memory and multiple accelerators. Long context is offered but degrades in effective use. The models occasionally mix languages or produce Chinese text when English is requested, particularly at small sizes. And the reasoning mode, while useful, can produce long traces that are billed as output 3.

A more consequential limit is political. Because the models are trained to comply with Chinese regulations, they decline or reframe questions on topics such as Taiwan, Tibet, Xinjiang and the 1989 protests. Developers outside China who deploy Qwen inherit that behaviour unless they fine-tune it away, and fine-tuning may not remove it reliably because the refusal is embedded in the training rather than in a removable filter 5.

Alibaba reports benchmark results for each release, and the numbers are vendor-reported. Qwen3-235B-A22B was presented as competitive with contemporary reasoning models on mathematics, science and coding, and the report lists results on MMLU, GPQA, AIME, LiveCodeBench and multilingual suites. Qwen3-Max was reported at or near the top of several leaderboards on release 2 6.

The comparisons require care. Alibaba has sometimes benchmarked against older versions of competitors' models, and the strongest proprietary systems are usually released after the open Qwen models, so a like-for-like comparison at any moment favours whoever released most recently. The report does publish its evaluation settings, which is better practice than most, but independent replication remains limited 6.

Independent leaderboards have generally confirmed that Qwen is strong for an open model, particularly in multilingual and coding tasks, while placing it below the best closed systems on hard reasoning. That is a respectable result and a different claim from the one implied by launch marketing. The gap matters less for the many users who value local control over marginal capability 3.

A specific measurement issue concerns the hybrid thinking mode. Scores depend heavily on whether the model was allowed to reason, how long the traces were and whether tools were available, and different evaluations have used different settings. A single number for "Qwen3-235B" therefore conceals a range of configurations, and comparisons that mix them are not meaningful 2.

Open-weight Qwen models are free to download from ModelScope and Hugging Face. Users accept a licence, which for Qwen3 is Apache 2.0 and therefore permits commercial use, modification and redistribution without a user threshold or a naming condition. That is a genuinely open licence, unlike the community licences used by some competitors 3.

Alibaba also sells API access through Model Studio, priced per thousand tokens in renminbi with separate rates for input and output and a free quota for new users. The proprietary models, including Qwen3-Max, are available only this way. Prices are lower than those charged by American labs for comparable tiers, which reflects both lower operating costs and a strategy of buying market share 7.

For Chinese enterprises the practical route is Alibaba Cloud, where Qwen is integrated with storage, databases and security services. Outside China, developers typically download the weights and run them on their own hardware or through a Western inference provider, because using the Chinese API raises data-residency and legal questions that many organisations prefer to avoid 1.

The consumer app is free, with a subscription for heavier use and access to the largest models. Alibaba has bundled Qwen into its existing products rather than charging separately, which is the same cross-subsidy strategy used by Google and Meta: the model is a feature of the platform, not a stand-alone business 4.

The most consistent criticism concerns censorship. Qwen models are trained to comply with Chinese content rules, and independent evaluations have found refusals or evasive answers on politically sensitive subjects that other models handle directly. Alibaba does not dispute the behaviour; it is a legal requirement in its home market. The dispute is about what it means for users elsewhere, who may not realise that a downloadable model carries the politics of its training environment 5.

A second criticism concerns data and provenance. Alibaba has not published its training corpus, and the 36-trillion-token figure cannot be verified. Some of that data is likely to include copyrighted material, as with every large model, and Chinese copyright law and enforcement differ from the American and European regimes. The company has not faced the kind of litigation that Anthropic and OpenAI have, which reflects jurisdiction as much as practice 6.

A third concerns benchmark presentation and the "open" narrative. Qwen's open models are genuinely open, but the company's strongest systems are proprietary, and launch comparisons often pair the open models against older competitors. Critics argue that this flatters the family; defenders note that Alibaba releases more weights, more often, than any Western lab of comparable size 2.

Geopolitics is the fourth theme. US export controls limit Chinese access to advanced chips, and Qwen's efficiency-focused designs — mixture-of-experts, distillation, small dense models — are partly a response to those constraints. American policymakers have debated restricting the diffusion of Chinese open models, while developers outside China have adopted them because they are free and capable. The family is therefore both a technical achievement and an object of policy conflict 1.

Whether Qwen leads or follows is disputed. Alibaba's benchmarks often show parity with or superiority over the best American models at the time of release, and independent leaderboards generally place Qwen near the top of the open-weight category while ranking it below the strongest closed systems. The disagreement is partly about timing, since open models are released later than the proprietary models they are compared with 6.

The safety question has two layers. Western critics argue that open weights plus Chinese content rules produce a model that is both censored and uncontrollable once downloaded. Alibaba and its defenders reply that the censorship is transparent and legally required, that open weights allow researchers to study and remove it, and that American open models carry their own biases. Neither side can claim the comparison is clean 5.

The "open source" claim is less contested than for Llama, because Apache 2.0 is an OSI-approved licence and Qwen3 uses it. The remaining argument is about the proprietary models: a family that releases its second-best models and keeps its best behind an API is open in a meaningful but limited sense, and the label "open-weight leader" describes that arrangement more accurately than "open source" 3.

What is not disputed is the effect on the market. Qwen and its derivatives became some of the most downloaded models on Hugging Face, and a large share of fine-tuned models published by Chinese developers are based on it. For a company whose core business is cloud computing, that adoption is the point: every derivative is a potential customer, and every benchmark result is an advertisement for Alibaba's infrastructure 1.