Alpha
All entriesContact
Model

Qwen

Publisher
Alibaba (Qwen team) (China)
Family
Qwen
Openness
open_weights
Licence
Apache License 2.0 (mainstream Qwen3 sizes; verify per checkpoint)
Context
32,768 native, 131,072 with YaRN (per checkpoint) / 262,144 native (per checkpoint)

Qwen is partly owned. The Apache-licensed mainstream Qwen3 sizes can be run, modified and self-hosted, they perform strongly (especially at coding), and self-hosted data stays on the deployer's infrastructure. Two things cap ownership at partial. First, the licence is per checkpoint: Apache-2.0 on mainstream Qwen3, but the legacy and large Tongyi Qianwen checkpoints (Qwen-72B, Qwen2.5-72B-Instruct) carry a >100M-MAU trigger and 'Built with Qwen' attribution, so verify the LICENSE on the exact checkpoint. Second, transparency is weak: China-aligned censorship is built in and training is closed. The hosted path carries a binding no-training-without-consent term (Alibaba Cloud Product Terms Sec 4.48) alongside a privacy policy that documents Singapore storage, SCC/DPF transfers and retention, but no document gives an exact retention period, a Model Studio storage region or an opt-out. Adopt with hard limits: keep politically sensitive and regulated workloads off it, confirm the per-checkpoint licence, and close the Article 55 gap yourself if a systemic-risk-scale variant goes on the EU market.

Do you really own it?
Partial
none·limited·partial·substantial·full
Analytical input: AOI C · 67.6/100
The four ownership factors

Floor-weighted, not averaged. Transparency is weak, and the weakest factor sets the ceiling, so the verdict stays partial however strong the rest.

1

Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?

Strong

The licence is per checkpoint. Mainstream Qwen3 (0.6B, 32B, 235B-A22B, Coder-480B-A35B-Instruct) is Apache-2.0 with unconditional commercial use and no MAU clause. The legacy and large Tongyi Qianwen checkpoints (Qwen-72B, Qwen2.5-72B-Instruct) carry a >100M-monthly-active-user licence trigger, a 'Built with Qwen' attribution requirement and export controls. Check the LICENSE file on the exact checkpoint before relying on Apache terms.

How this scores (AOI sub-dimensions)
Openness3/5how much is released - weights, data, code, licence - and how freelyOpen-weights tier: weights are downloadable under Apache-2.0 with a good model card.
Legal3/5how permissive and clean the licence is for real commercial useThe permissive Apache-2.0 licence on the mainstream sizes is better than a restricted community licence.
2

TransparencyDo you know what it is: weights, training, behaviour, and legible terms?

Weak

The weights are held but cannot be inspected: training data and code are closed and China-aligned topic censorship is built into behaviour, so sensitive-topic coverage is distorted in ways that cannot be audited. On the hosted path, the binding Alibaba Cloud Product Terms (Sec 4.48, updated 19 Oct 2026) say Alibaba Cloud 'will not use your Member Content to develop or improve the models on Model Studio, unless you separately provide your consent'. The FAQ repeats a softer 'never uses your data for model training'. The Terms state no retention period, storage region or opt-out mechanism.

How this scores (AOI sub-dimensions)
Provenance3/5how well we can trace and verify what went into the modelThe Qwen org on Hugging Face is verified, the models are also officially published on ModelScope, and checkpoints are safetensors with checksums and no canonical-org malicious incident (checklist 5/8).
Governance3/5how accountable and well-documented the publisher isAlibaba's Qwen team is an active, accountable publisher with a rapid release cadence and a verified presence on two hubs.
3

ReliabilityIs it reliable and good enough for the job?

Moderate

Qwen3 runs dependably with 32K native context (128K with YaRN; 256K native on Coder) and documented thinking modes. Safety tuning is lighter than at Western frontier labs and no companion guard model exists, so the deployer supplies that layer.

How this scores (AOI sub-dimensions)
Operational5/5how practical it is to run, serve and maintain in productionDistribution is dual, on Hugging Face and ModelScope.
Safety3/5whether misuse risks are evaluated and guardrails are providedInstruct variants are safety-tuned and withstand casual jailbreaks.
4

Doesn't extract your dataDoes running it keep your knowledge and data yours?

Strong

Self-hosted, the Apache-licensed weights run on the deployer's own infrastructure with no clawback or telemetry, so data stays with the deployer. On Alibaba's hosted Qwen, the binding Product Terms (Sec 4.48) commit that Alibaba Cloud will not use Member Content to develop or improve its Model Studio models without separate consent. They treat Input and Output as the customer's content, processed on the customer's behalf. The FAQ adds AES-256 encryption of transmitted data. The general privacy policy documents Singapore storage (Philippines addendum), SCC/DPF cross-border transfers and 'ongoing legitimate business need' retention. No document states an exact retention period, a storage region for Model Studio content or an opt-out mechanism, and the Terms let Alibaba Cloud remove stored data from time to time.

How this scores
Not a scored AOI dimension. For a self-hosted model, data-control is a structural property of running the weights yourself, strong by default unless the model phones home or the licence claws back rights. For a hosted API this factor is the retention + train-on-inputs + residency read, scored from the binding terms.

How the AOI score is computed

The seven dimensions above, each scored 0 to 5, weighted and summed to the 0 to 100 headline. The score is the analytical input behind the ownership verdict, not the verdict itself.

DimensionScoreWeightPoints
Openness3/50.1810.8
Provenance3/50.169.6
Legal3/50.169.6
Safety3/50.169.6
Performance4/50.1411.2
Operational5/50.1212.0
Governance3/50.084.8
HeadlineC · 67.6/100

Grade ceiling: a hard flag (🚩 EU AI Act systemic-risk variants (Qwen3-235B, Qwen3-Coder-480B) likely exceed 1e25 FLOPs with no Article 55 documentation) caps the grade below the raw band. See the classification matrices.

Dossier coverageAssess 87%Implement 91%Use 67%Support 50%How complete our four-domain documentation is, a measure of our coverage, not of the model. Each domain links to its page.

Sources

Every rating traces to a primary document. Read means the text was verified; unverified means it is known to exist but has not yet been read.

DocumentWhat it grounds
Model cardread2026-11-19
Qwen3 checkpoints are hosted on the verified Qwen org on Hugging Face in safetensors with checksums.
Licenceread2026-11-19
Mainstream Qwen3 checkpoints are released under Apache-2.0 with unconditional commercial use and no monthly-active-user clause.
Licenceread2026-11-19
Legacy and large Tongyi Qianwen checkpoints (Qwen-72B, Qwen2.5-72B-Instruct) are not Apache-2.0.
Technical_reportread2026-11-19
The Qwen3 Technical Report (arXiv 2505.09388) documents the Qwen3 family, thinking/non-thinking modes, thinking budget and multilingual coverage (~119 languages).
Privacy Policyread2026-11-19
The general Alibaba Cloud International Website Privacy Policy (Last Updated 4 Sep 2026) ties retention to "ongoing legitimate business need" (Sec F).
Documentationread2026-11-19
The Alibaba Cloud Model Studio FAQ states: "Alibaba Cloud strictly protects data privacy and never uses your data for model training." All data transmitted "is encrypted with AES-256".
Terms of serviceread2026-11-19
Alibaba Cloud International Website Product Terms of Service (Last Updated 19 Oct 2026), Sec 4.48 Model Studio, treat Input and Output as "Member Content" processed "for your purposes and on your behalf".
Model cardread2026-11-19
Qwen is also officially published on Alibaba's ModelScope hub (dual distribution).
Third-party analysisunverified2026-07-25
No public EU AI Act training-content summary, copyright policy, or Article 55 documentation for the systemic-risk large variants.
Third-party analysisunverified2026-07-25
Qwen instruct variants are safety-tuned but with lighter alignment coverage than Western frontier labs.
Third-party analysisunverified2026-07-25
Qwen models exhibit topic censorship aligned with Chinese content rules.
Third-party analysisunverified2026-07-25
Qwen3-Coder is among the strongest open coding models on independent evaluation.
Third-party analysisunverified2026-07-25
Qwen is broadly supported across serving stacks (vLLM, llama.cpp, Ollama, TGI, MLX) with many community quants.