The Free Qwen Stack for Developers
Search results for a “free Qwen API” often drop you on the international Model Studio site, a Qwen Code subscription menu, or a local-install guide. If you need to call current Qwen models from mainland China without standing up a GPU box or buying a plan, those are the wrong first pages. The path that is still free, official, and completable after 实名认证 is Alibaba Cloud Model Studio on the China site, in 华北2 / China (Beijing).
This page is the orientation for that path. It tells you what the stack includes, the rules every tutorial in this cluster assumes, and which of the eight written guides to open first. It is not a three-model bake-off, and it is not a recap of eight openings.
Here is the short version:
- Work on the Alibaba Cloud China site (aliyun.com). Finish 实名认证, activate Model Studio, and create an ordinary API key in China (Beijing). You can finish that path without a foreign legal entity, without a Token Plan, and without adding funds first.
- Each eligible model ID gets its own new-user grant. For the text models in this cluster, the China-site Beijing price table still lists 1,000,000 tokens for 90 days, with input and output (including thinking) sharing one pool. For
qwen-image-3.0, the grant is 10 images, and inputs and outputs share that pool. You cannot fold two model IDs into one pool, and exhaustingqwen4-maxdoes not automatically start spendingqwen3.8-flash. - Point your client at the Beijing OpenAI-compatible host,
https://{WorkspaceId}.cn-beijing.maas.aliyuncs.com/compatible-mode/v1, and send the unsuffixed model ID (qwen4-max, notqwen4-max-preview;qwen3.8-max, notqwen3.8-max-0902). Dedicated Token Plan and Coding Plan keys are the wrong credential for this allowance. - Before any call, enable 免费额度用完即停 for each model you will use. The English name is Free Quota Only. The switch is off by default. Leave it off, and an exhausted grant flips to the published list price. Enable it, and the official free-quota page documents a hard stop: HTTP 403, error code
AllocationQuota.FreeTierOnly, and no further charge. Do not top up “just in case” before you have used the grant and have decided you want paid traffic. - Official pages for these rules, prices, and listing dates were opened on 2026-11-17 (Asia/Shanghai). This site has no live Beijing API key on record for this batch. The eight tutorials were written from those docs. They are labeled Not yet verified, and nothing on this page is a measured run.
What does this free Qwen stack actually include?
On this site, “the free Qwen stack” is a shorthand for one credential path plus the jobs you can do on it. Alibaba does not sell a product under that name. The path is the China-site Beijing new-user grant, a regular Model Studio key, and the region’s compatibility host.
That is a different offer from the pages that usually rank for nearby queries. The international documentation (the alibabacloud.com and Model Studio English console docs) describes a Singapore grant for the International deployment scope. The China-site documentation, including the 新人免费额度 page opened for this article, says the opposite geography: only models in 华北2 (Beijing) have a free quota, and other regions on that site have none. Stay on the China site if that is where you can finish signup and verification. Do not paste a Singapore host in front of a Beijing key and expect the China-site grant to apply.
This path is also not Qwen Code’s Plus or Pro sign-in, and it is not a Token Plan. Those are paid or subscription routes with their own keys and hosts. If the key belongs to a Token Plan or a Coding Plan, the official quota page says the new-user allowance is not what gets deducted; those dedicated keys bill on their own path. For the Beijing grant, use a general Model Studio key.
What you can do on that path, in this cluster, is eight specific jobs: a first tool-calling call on qwen4-max, a first hosted call on qwen3.8-27b, a generate-then-edit image job on the 10-image grant, a decision among three text-model grants, a three-value swap of an existing OpenAI-style client, image and short-video understanding on qwen4-max, a reasoning_effort sweep, and a checkable review through Alibaba’s Open Code Review CLI.
What should you set up before you open any tutorial?
The eight guides repeat a few facts because each one has to stand alone. You only have to get them right once.
Create the account and the key on aliyun.com, after 实名认证, in China (Beijing). Official quota rules say a grant is not issued until that verification is done, and that a first Model Studio login activates the service without a separate product purchase. The console, not this page, is where you confirm that a given model actually received a grant, how many tokens or images remain, and when the window ends.
A grant lasts 90 days. The start date is not automatically “today”: official docs pick the latest of three events — the day you activated Model Studio, the day the model was listed, or the day a request for that model was approved. Listing dates on the China-site 模型上架与更新 page, re-fetched on 2026-11-17 (Asia/Shanghai), and the date 90 days after each listing, are:
| Model ID | Official listing date | 90 days after that listing date |
|---|---|---|
qwen3.8-max |
2026-08-02 | 2026-10-31 |
qwen-image-3.0 |
2026-08-04 | 2026-11-02 |
qwen3.8-27b |
2026-08-17 | 2026-11-15 |
qwen3.8-flash |
2026-08-26 | 2026-11-24 |
qwen4-max |
2026-11-16 | 2027-02-14 |
Those end dates are only examples for an account that was already open on the listing day. As of 17 November 2026, that would mean the qwen3.8-max, qwen-image-3.0, and qwen3.8-27b windows may already have closed, the Flash window may be close, and the qwen4-max window is still early. A first-time activation on or after 17 November 2026 is a later start than those listing dates, so the official “whichever is later” rule can still open a fresh 90-day window. This site has not opened a console to confirm that a brand-new account still receives every grant the price table lists. Trust the expiry printed next to the model.
The China-site 模型调用价格 table, opened the same day, still attaches Beijing grants to the unsuffixed IDs above: 1,000,000 tokens for the four text models, and 10 shared input-and-output images for qwen-image-3.0. The snapshot qwen3.8-max-0902 is on the same table at the qwen3.8-max list price and is marked as having no free quota. qwen4-max-preview is a Token Plan model, listed on 2026-11-11, and is the wrong ID for a regular free-quota key.
After a grant is exhausted or expired, continued real-time calls bill at list price unless the stop switch is on. Beijing list prices on that same table, for the first pricing tier (input of at most 1 million tokens in one request), are CNY 0.8 / 2.7 per million input / output tokens for qwen3.8-flash, CNY 3 / 12 for qwen3.8-27b, CNY 10 / 30 for qwen4-max, and CNY 12 / 36 for qwen3.8-max. For qwen-image-3.0, input is CNY 0.02 per image and output is CNY 0.18 per successful image at both 1K and 2K. Output token prices include thinking plus the visible answer. This page is not the place to pick among the three larger text models; that comparison lives in Qwen3.8-Flash vs Qwen3.8-Max vs Qwen4-Max: Which Beijing Free Quota to Use.
Official quota docs also limit what the grant will cover. It applies to real-time inference. It does not cover batch calls, fine-tuning, deployment, custom models, PAI-DSW, or OSS storage and request fees. Usage by the primary Alibaba Cloud account and by its RAM users comes out of the same per-model remainder.
Turn the stop switch on in the Beijing console, on a computer, before the first call. The 新人免费额度 page says the control is off unless you enable it, that the mobile console cannot toggle it, and that a change can take a short time to apply. The same page warns that production traffic which must keep running after the grant ends should not rely on this halt. For the tutorials in this cluster, the halt is the point: you are testing on the grant, not sliding into pay-as-you-go.
Which tutorial should you open first?
The eight guides are written drafts in this cluster, linked below with site-relative paths. They are not claimed as already-live public URLs, and none of them yet has a Beijing run in this site’s records. Start from the job you actually have.
If you want a first checkable call on the current flagship, open Call Qwen4-Max Through the OpenAI-Compatible API and Finish a Tool-Calling Task. It points the official OpenAI SDK at qwen4-max and plans a package.json audit whose report you can check locally, without trusting a fluent reply. It also separates qwen4-max from the Token Plan preview ID.
If you came from the Hugging Face card or a self-host guide for the open 27B model, open Call the Official Hosted Qwen3.8-27B API on Alibaba Model Studio. The hosted ID is qwen3.8-27b, not Qwen/Qwen3.8-27B. The planned job extracts five fields from a pytest failure and checks them on your machine.
If you already have a working OpenAI, Responses, or Anthropic client, open Switch an Existing OpenAI Client to Qwen4-Max Without Rewriting It. You keep the request code you already have and replace only the credential, the base URL, and the model ID. That tutorial is about surviving the swap — including parameters that no longer mean what they did — not about writing a first script from scratch.
If the question is which of three Beijing text grants to spend, open Qwen3.8-Flash vs Qwen3.8-Max vs Qwen4-Max: Which Beijing Free Quota to Use. That page is the comparison. This hub will not repeat its same-task design or fill in a quality winner. Until a live run exists, it chooses from official positioning, list prices, and whether each grant is still on the account.
If you need a still image rather than a text answer, open Call Qwen Image 3.0 for One Product Shot and One Edit on the 10-Image Free Quota. It budgets a single generated mug, then a recolor of that same shot, and it counts those calls as three images on the shared 10-image pool.
If you already have a picture or a short clip and want a text description you can check, open Use Qwen4-Max to Understand One Image and One Short Video on the Beijing Free Quota. The model ID stays qwen4-max; you do not switch to a VL-only name. The output is text. The model is not documented to understand a video soundtrack, and this is not an image generator.
If thinking mode is burning the grant before you see an answer, open Control Qwen4-Max Thinking With reasoning_effort. It stays on one model and one grant, and it plans a four-level sweep (none, low, high, xhigh) against a fictional yard invoice whose correct total is already known.
If you want a review CLI rather than a chat completion you parse yourself, open Review a Checkable Python Fixture with Open Code Review and Qwen4-Max. The CLI is Alibaba’s open-source Open Code Review tool (Apache License 2.0). The CLI is not a paid review plan; qwen4-max on the Beijing grant is what spends tokens. The fixture is a small original allocator with planted issues you can score against a checklist.
What has this site not verified yet?
None of the eight tutorials, and not this hub, has been run against a live Beijing Model Studio key in this site’s records. Official documentation was read on 2026-11-17 (Asia/Shanghai). Example scripts next to the tutorials were syntax-checked, and some local fixtures or dry-run paths were exercised offline. That is not a substitute for an API transcript, a console screenshot, a latency number, or a token count the service itself returned.
Until a run exists, treat response shapes, error bodies, and “what the model will say” as documented, not observed. Do not read this cluster as proof that the stack works end to end. When a Beijing run is captured, the individual tutorials are the place to record the date, the package versions, and the output — not a later claim on this page that “the stack was verified.”
What is left out of this cluster?
Two local Ollama topics were in the original batch and were deferred. This writing environment does not have the roughly 18 GB of memory or VRAM needed to verify those installs. They are out of scope here until they can be checked on suitable hardware. This page does not sketch unverified install steps for them.
The cluster also does not walk through Token Plan, Coding Plan, Qwen Cloud’s international hosts, or the Singapore grant described on the international site. It does not treat batch inference, fine-tuning, or deployment as part of the free path; official quota docs exclude those charges from the grant. The preview ID qwen4-max-preview is also out of scope; official pages restrict it to Token Plan access.
What should you do next?
Pick the tutorial that matches the job you have, not the one with the newest model name. On aliyun.com, finish 实名认证, open Model Studio in China (Beijing), and confirm the grant and the expiry in the console. Enable 免费额度用完即停 / Free Quota Only for the model IDs you will actually call, wait for that setting to apply, and stay on a regular Beijing key. You do not need to add funds first.
If you only need one starting point and you do not already have a client to migrate, begin with the qwen4-max tool-calling guide. If you already have an OpenAI-style client, begin with the swap guide instead.
Sources
Official China-site pages opened or re-fetched on 2026-11-17 (Asia/Shanghai) for the shared rules, prices, and listing dates on this page:
- 新人免费额度: Beijing-only grants on the China site, 90-day clock, per-model pools, 实名认证, 用完即停 default and HTTP 403 /
AllocationQuota.FreeTierOnly, coverage limits, Token Plan / Coding Plan keys - 模型调用价格: Beijing list prices and grant sizes for the model IDs named above, including
qwen3.8-max-0902marked as having no free quota - 模型上架与更新: listing dates used in the clock table
- Qwen4-Max and Qwen4-Max-Preview: production versus Token Plan preview IDs
The eight tutorials in this cluster, with the site-relative paths used above:
- Call Qwen4-Max Through the OpenAI-Compatible API and Finish a Tool-Calling Task
- Call the Official Hosted Qwen3.8-27B API on Alibaba Model Studio
- Call Qwen Image 3.0 for One Product Shot and One Edit on the 10-Image Free Quota
- Qwen3.8-Flash vs Qwen3.8-Max vs Qwen4-Max: Which Beijing Free Quota to Use
- Switch an Existing OpenAI Client to Qwen4-Max Without Rewriting It
- Use Qwen4-Max to Understand One Image and One Short Video on the Beijing Free Quota
- Control Qwen4-Max Thinking With reasoning_effort
- Review a Checkable Python Fixture with Open Code Review and Qwen4-Max
These pages were read only to see how the topic already appears in search results. Their setup advice was not used as a spine for this hub:
- Free quota for new users (international / Singapore documentation)
- Qwen Code Quickstart
- Qwen4-Max Leaves Preview: What the Price Card Actually Says (OrcaRouter)
- Alibaba Cloud Model Studio new-user free quota (Free LLM Credits)
- Alibaba Cloud Model Studio’s 90-Day Free Quota: Activation, Billing and Spending Protection (PuppyIP)
Title tag: The Free Qwen Stack for Developers on Beijing Model Studio
Meta description: A map of the China-site Beijing Model Studio free-quota path for Qwen, the stop-when-exhausted rules, and eight developer tutorials. Written from official docs; not yet run.
Slug: free-qwen-stack-for-developers