Qwen is a family of large language models and multimodal models developed by the Qwen Team at Alibaba Cloud. Alibaba introduced the broader Tongyi Qianwen product in April 2023 and released the first Qwen model weights later that year.[1][2] The family now includes general-purpose text models, vision-language systems, audio and image models, coding and mathematics variants, embedding models, and both dense and mixture-of-experts architectures.[21]

Qwen is not a single model with one fixed license or access method. Some checkpoints can be downloaded under the Apache License 2.0, earlier or larger checkpoints use Qwen-specific licenses, and several Max, Plus, and other hosted models are available only through Alibaba services. This article therefore uses open-weight for downloadable parameters and reserves open source for a system whose released materials meet a fuller definition that includes sufficient training-data information and code.[3]

Scope and naming

Alibaba announced Tongyi Qianwen on April 11, 2023 as a large language model intended for integration across its businesses and for customer customization.[1] "Tongyi Qianwen" remains a broader product and service name, while "Qwen" is the model-family name used in technical reports, model repositories, and international documentation. The separate Tongyi Qianwen article covers that broader product identity.

This overview summarizes the main language-model generations and their access conditions. It does not replace the more detailed articles on Qwen2, Qwen2.5, Qwen3, Qwen3.5, Qwen3.6, Qwen3-Max, or Qwen3.7-Max. Specialized branches including Qwen-VL, Qwen2-Audio, Qwen Image, Qwen3-Omni, Qwen3-Coder, Qwen2-Math, QwQ, and Qwen3-Embedding also retain their own scopes.

Development history

Initial Qwen releases

The September 2023 Qwen technical report described Qwen as a series containing base pretrained models, chat models aligned with human feedback, and code- and mathematics-specialized variants. The first report evaluated Qwen-7B and Qwen-14B and documented decoder-only Transformer models, tokenizer and training choices, supervised fine-tuning, and reinforcement learning from human feedback.[2] The downloadable Qwen-7B repository's custom Tongyi Qianwen License recorded an August 3, 2023 release date. That license permitted many uses but required separate permission for commercial services above a stated monthly-user threshold and restricted use of the materials or their output to improve a different large language model.[4] It was therefore a weight release, not an Apache-licensed release.

Qwen1.5

Qwen1.5 was announced on February 4, 2024. The initial release ranged from 0.5 billion to 72 billion parameters, with additional mixture-of-experts and 110-billion-parameter models released separately. The Qwen Team reported a 32,768-token context length for the initial language-model lineup and integration with the Hugging Face Transformers library.[5] License terms were not uniform across every size, so a generation-level statement such as "all Qwen1.5 models are Apache-licensed" would be inaccurate.

Qwen2

Qwen2 was announced on June 7, 2024 with four dense sizes, 0.5B, 1.5B, 7B, and 72B, plus the Qwen2-57B-A14B mixture-of-experts model. The release added training support for 27 languages beyond Chinese and English. The announcement offered context lengths of up to 128,000 tokens for the 7B and 72B models, not for every size.[6] The technical report says most Qwen2 models were pretrained on more than 7 trillion tokens, while the 0.5B model and the mixture-of-experts model followed different schedules.[7]

Qwen2 also illustrates why licenses must be checked checkpoint by checkpoint. The Qwen2-7B repository is marked Apache License 2.0, while Qwen2-72B carries the custom Tongyi Qianwen license.[8][9] Parameter count, context length, instruction tuning, and license are separate properties; a statement that is true for one Qwen2 checkpoint does not automatically apply to another.

Qwen2.5

Qwen2.5 was announced on September 19, 2024 after pretraining on up to 18 trillion tokens. The language-model lineup included 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B sizes, alongside Coder and Math branches.[10] Its technical report records more than one million supervised fine-tuning examples and over 150,000 preference pairs, but it does not publish the complete training corpus.[11]

The same report gives model-specific context and license information:

Qwen2.5 size	Context / maximum generation in the report	Weight license
0.5B and 1.5B	32K / 8K	Apache License 2.0
3B	32K / 8K	Qwen Research License
7B, 14B, and 32B	128K / 8K	Apache License 2.0
72B	128K / 8K	Qwen License

These values come from Table 1 of the Qwen2.5 technical report.[11] They correct two common overgeneralizations: the smaller models did not all have a 128K context in that report, and the full lineup did not share one license.

Qwen3

Qwen3 was announced on April 29, 2025. The released family contained six dense models, from 0.6B to 32B, and two mixture-of-experts models: Qwen3-30B-A3B and Qwen3-235B-A22B. The Qwen Team released the eight models under the Apache License 2.0.[12] Its technical report describes training on 36 trillion tokens spanning 119 languages and dialects.[13] Official Qwen sources disagree on one context value: the release post lists 32K for the 0.6B, 1.7B, and 4B models and 128K for the 8B, 14B, 32B, and two mixture-of-experts models, while Table 1 of the technical report lists 128K for Qwen3-4B. This overview therefore does not assign Qwen3-4B one unqualified context figure.[12][13]

Qwen3 joined a thinking mode for multi-step reasoning and a non-thinking mode for faster direct responses in the same post-trained model. The release also exposed a configurable thinking budget.[12][13] These are inference modes, not guarantees that a longer reasoning trace will be correct. They also should not be confused with Qwen3-Max, a separately hosted branch. Alibaba described the September 2025 Qwen3-Max release as a proprietary API model with more than one trillion parameters and 36 trillion training tokens; its weights were not released.[14]

Qwen3.5 and Qwen3.6

Qwen3.5 was announced on February 15, 2026 with Qwen3.5-397B-A17B as its initial open-weight model. The name indicates 397 billion total parameters and 17 billion activated for a token. The model combined a vision encoder with a language model using Gated DeltaNet layers, full-attention layers, and sparse experts. The model card lists 201 supported languages and dialects, an Apache License 2.0 release, a native context length of 262,144 tokens, and optional extension to 1,010,000 tokens.[15][16] The hosted Qwen3.5-Plus service offered a one-million-token context by default. The announcement said that a technical report would follow; as of July 28, 2026, no official Qwen3.5 report disclosing a complete training-token total or full corpus had been published.

The first open-weight Qwen3.6 model, Qwen3.6-35B-A3B, was announced on April 15, 2026. It is a native vision-language mixture-of-experts model with 35 billion total and 3 billion activated parameters. Qwen3.6-27B, a dense native vision-language model, followed on April 22.[17][18] Both model repositories use the Apache License 2.0 and list a native 262,144-token context length with optional extension to 1,010,000 tokens.[19][20] The model cards warn that static YaRN scaling can reduce performance on shorter inputs, so the extrapolated maximum should not be read as a measured guarantee for every task or deployment.

Hosted Qwen3.7 and Qwen3.8 services

The Qwen site records the release of Qwen3.7-Max on May 20, 2026 and Qwen3.7-Plus on June 1, with the latter adding multimodal input. Alibaba Cloud's release log lists their global service availability on May 21 and June 2, respectively, and identifies the dated snapshots as qwen3.7-max-2026-05-20 and qwen3.7-plus-2026-05-26. The same log records the global release of qwen3.7-max-2026-06-08 on June 10; that snapshot added visual input, whereas the May 20 Max snapshot was text-only.[21][23] Alibaba Cloud's July 2026 model documentation lists the hosted Max and Plus models with context limits up to one million tokens.[22] No official downloadable weights, parameter counts, full training-data description, or technical report for Qwen3.7-Max or Qwen3.7-Plus had been released by the cutoff date.

Alibaba announced Qwen3.8-Max-Preview on July 20, 2026 and made it available through its Token Plan, Qoder, and QoderWork products. Alibaba described it as a 2.4-trillion-parameter model and said Qwen3.8-Max would become open weight "soon."[24] As of July 28, it was still a preview service rather than a documented open-weight release. No downloadable weights, weight license, technical report, model card, training-token count, or fixed weight-release date had been published. The 2.4-trillion figure is an attributed company disclosure, and "soon" is a stated plan rather than a release date.

Architecture and training disclosure

The early Qwen language models were decoder-only Transformer systems trained with next-token prediction. The reports describe components including rotary position embeddings, RMS normalization, SwiGLU feed-forward networks, and grouped-query attention in later generations.[2][7] Qwen2.5 retained dense Transformer models and also documented a fine-grained mixture-of-experts design. Qwen3 used dense and sparse-expert variants, with 128 total experts and eight activated experts in each of its two mixture-of-experts configurations.[11][13]

Qwen3.5 and Qwen3.6 changed the backbone more substantially. Their released model cards describe repeating groups of Gated DeltaNet linear-attention layers followed by a full-attention layer. Some checkpoints are dense and others route tokens through a subset of experts. They also include a vision encoder rather than requiring a separately named VL checkpoint for image input.[16][19][20] Architecture labels do not make parameter counts interchangeable: for example, "35B-A3B" reports both total and active parameters, while "27B" describes a dense model.

The Qwen reports disclose aggregate token counts, high-level data domains, filtering stages, synthetic-data use, supervised fine-tuning, preference optimization, and reinforcement-learning procedures. They do not release a complete, itemized training corpus with provenance for every example or all code needed to reproduce training.[2][7][11][13] This distinction matters for both scientific reproducibility and license terminology.

Weights, licenses, and access

Qwen releases fall into three broad access groups:

Custom-licensed open weights. The original Qwen-7B and some later large checkpoints use Qwen-specific terms.[4][9][11]
Apache-licensed open weights. Examples include Qwen2-7B, most but not all Qwen2.5 sizes, all eight initial Qwen3 models, and the cited Qwen3.5 and Qwen3.6 checkpoints.[8][11][12][16][19][20]
Hosted proprietary models. Max and Plus services can be available through Qwen Chat or Alibaba Cloud without downloadable weights. Qwen3-Max and Qwen3.7 are examples.[14][22]

An Apache License 2.0 label applies to the covered released materials, but a license on weights does not by itself disclose the data and process used to derive them. The Open Source Initiative's Open Source AI Definition requires the ability to use, study, modify, and share a system, plus the preferred form for modification, including detailed data information, complete training and inference code, and parameters.[3] Because Qwen's public reports provide high-level descriptions rather than the complete training corpus and pipeline, "Apache-licensed open-weight model" is more precise than claiming that every released Qwen system is fully open source.

Hosted services and APIs

Alibaba Cloud Model Studio serves Qwen and third-party models across multiple regions. Its July 13, 2026 text-generation API reference lists four interface families: OpenAI-compatible Chat Completions, OpenAI-compatible Responses, Anthropic-compatible Messages, and the native DashScope interface.[25] Model and feature availability varies by region and by exact model identifier, and this overview does not inventory every hosted service alias.[22][25]

Unversioned service aliases can change as Alibaba updates a hosted model. Where Model Studio offers a dated snapshot, applications that require reproducibility should record the exact identifier, region, request parameters, and access date. A context-window limit in service documentation is an input-and-output capacity boundary, not evidence that a model will retrieve or reason correctly over every position in a long prompt.[22][23]

Evaluation and limitations

Qwen technical reports and model cards contain extensive vendor-run benchmark comparisons, but scores across generations are not automatically comparable. Prompt templates, sampling settings, context limits, tool scaffolds, benchmark revisions, and judging models can all change. Results should be attributed to the documented setup rather than converted into a timeless ranking.[7][11][13][16]

Independent studies provide narrower evidence. QualBench, published at EMNLP 2025, evaluated 17,316 Chinese professional-qualification questions across six domains. In its reported setup, Qwen2.5-7B-Instruct scored 75.26 percent overall, compared with 61.61 percent for GPT-4o, while the authors also documented domain variation, answer-shuffling tests, and weak performance on a question-completion probe.[26] This supports performance on that benchmark, not a conclusion that the smaller model is generally superior.

MuBench, published in Findings of ACL 2026, evaluated multilingual understanding in 61 languages. It found Qwen2.5 and Qwen3 models comparatively strong and stable among the tested open models, but still measured persistent losses relative to English, especially for lower-resource languages. Those gaps did not consistently disappear as model size increased.[27] A vendor statement that a model "supports" many languages therefore describes coverage, not equal accuracy across languages or tasks.

Other limitations are disclosure and deployment specific. Optional million-token extrapolation requires framework and scaling choices that can affect shorter-context behavior.[16][19][20] Hosted aliases can change, and proprietary models may expose fewer architecture or training details than open-weight checkpoints.[22][23] None of the cited reports establishes that Qwen outputs are uniformly factual, that benchmark results transfer to a particular application, or that a longer context or thinking mode eliminates errors.

See also

References

Editing: Qwen