Draft
Qwen3.8-Flash vs Qwen3.8-Max vs Qwen4-Max: Which Beijing Free Quota to Use
Task
What you will complete
Send the same changelog fixture to qwen3.8-flash, qwen3.8-max, and qwen4-max on the Beijing OpenAI-compatible endpoint, with thinking off, stay on each model's own free quota with 免费额度用完即停 / Free Quota Only turned on, and score each JSON reply against a fixed expected report.
Success criteria
- compare_three_quotas.py sends the same changelog.txt to qwen3.8-flash, qwen3.8-max, and qwen4-max, with thinking turned off.
- Each response is scored by the local checker against examples/expected.json (identity fields, seven breaking changes, two compatible additions, and no invented changes).
- A results table can be filled from examples/results/summary.json with request IDs, usage, and checker scores. Until a live run exists, those cells stay empty on purpose.
What this is not
Not a news roundup, not a quality ranking written in advance of a run, and not a guide to snapshot IDs, Token Plan preview IDs, batch inference, or the Singapore grant. The official cards can be cited; the quality ranking is unknown until the same-task run exists.
Prerequisites
- A China-site (aliyun.com) Beijing Model Studio account, with a remaining new-user grant visible in the console for each of the three model IDs you plan to call.
- 免费额度用完即停 (Free Quota Only) turned on for each of those model IDs before any call.
- Your workspace ID for the Beijing compatible-mode/v1 base URL.
- Python and the official openai package.
Environment
- DASHSCOPE_API_KEY
- Beijing Model Studio API key
- QWEN_BASE_URL
- https://{WorkspaceId}.cn-beijing.maas.aliyuncs.com/compatible-mode/v1
On Alibaba Cloud Model Studio’s China site in Beijing, qwen3.8-flash, qwen3.8-max, and qwen4-max each carry a 1,000,000-token new-user grant. Those pools are separate. Spending one does not deplete the others, and the platform will not switch models for you when a grant runs out. The useful question is which grant to spend on a given job — and whether that grant is still on the account.
Prices, listing dates, context limits, and capability flags below come from official Model Studio pages read on 2026-11-17 (Asia/Shanghai). This site has not called any of the three models for this comparison. There is no output table to rank, no latency number, and no token count from a real request.
A documented way to choose, until a live run fills the gap:
- Choose
qwen3.8-flashwhen the job is high-volume or exploratory and you can check the result yourself. The China-site text-generation guide names this ID as the cost-first pick. Beijing list prices, once the grant is gone, are CNY 0.8 per million input tokens and CNY 2.7 per million output tokens (thinking plus the answer). Input is 12.5 times lower thanqwen4-maxand 15 times lower thanqwen3.8-max. - Choose
qwen4-maxwhen the job needs the current flagship. The same official guide lists it first for harder reasoning and understanding. After the grant, Beijing list prices are CNY 10 / 30 per million input / output tokens — a sixth belowqwen3.8-maxat CNY 12 / 36. The official summary card gives this model 262,144 tokens for thinking plus the answer. The 3.8 model pages still cap the answer at 131,072 in thinking mode, with a separate 262,144-token chain-of-thought limit. - Choose
qwen3.8-maxonly if its grant still appears in the console and you have a reason to spend that pool: a second flagship allowance, a behavior you already measured on this generation, or, after the grant, a Batch File job (this ID lists a 50% Batch File rate;qwen4-maxlists batch inference as unsupported). Do not treat it as the current first-choice Max model. The official guide now files it under older models. - Open the console first. Official quota rules time the 90 days from Model Studio activation, the model’s listing date, or approval, and they use whichever of those dates is later. Official listing dates are 2026-08-02 for
qwen3.8-max, 2026-08-26 forqwen3.8-flash, and 2026-11-16 forqwen4-max. If the account was already active on those dates, the first of those clocks may already have expired as of 17 November 2026, and the Flash clock may be close. Read the date printed next to each model in the console rather than counting 90 days from this article.
Turn on 免费额度用完即停 (Free Quota Only in the English console) for each of the three IDs you might call. The switch is off by default. Leave it off, and a grant that runs out continues as pay-as-you-go at the prices above. Turn it on, and official docs say the next request is rejected instead, with HTTP 403 and the error code AllocationQuota.FreeTierOnly. The same page warns that this halt is intentional: the service stops when the tokens are gone, which is what a zero-spend test needs.
This comparison has no live outputs yet
The companion script is examples/compare_three_quotas.py. It passed python -m py_compile. It has not been executed against Model Studio, because this environment has no API key.
The official cards below are dated and attributed. The quality ranking is not. When a real run exists, the empty results table in this page will be replaced with that run’s date, package version, request IDs, token usage, and checker scores.
What same task will decide the quality question?
The same changelog goes to all three models, through the same Beijing OpenAI-compatible endpoint, with thinking turned off. A local checker then scores the JSON against a fixed expected report. Nothing in the score depends on how fluent the prose looks.
The fixture is examples/changelog.txt. It describes a fictional Quarrypond Courier Tracking API migration from v1 to v2, with a public sunset date of 2027-03-15. The contract hides seven breaking changes and two compatible additions in ordinary documentation language, then lists three items that did not change. The model has to return one JSON object:
- Identity fields:
api,from_version,to_version,sunset_date. - Seven breaking changes, addressed by path:
eventrenamed,parcel.weight_kgtype change,customer_emailremoved,statusenum change,pagereplaced bycursor, replay status changed from 200 to 409, andoccurred_atnewly required. - Two compatible additions: optional
retry_after_ms, optionalparcel.dimensions_cm. - No invented breaking changes, and no report of
tracking_id, theX-Quarrypond-Signatureheader, or the 120-request-per-minute limit as a change.
This is not a weather demo, and it is not the package.json audit or pytest extraction used in earlier tutorials in this series. The point is a result you can mark right or wrong without trusting a summary.
Thinking is disabled for the scored call (enable_thinking: false in extra_body). Official deep-thinking docs list all three IDs as hybrid thinking with thinking on by default. A default call will spend chain-of-thought tokens from the same 1,000,000-token pool, billed as output, before you ever see message.content. The comparison script turns that off so the grant pays for the JSON the checker can read. If you later rerun with thinking left on, record both the scores and the extra completion tokens. Do not treat those two setups as the same measurement.
What you need in order to run it
Work from mainland China on the Alibaba Cloud China site, aliyun.com. Finish 实名认证, open the 百炼 console in 华北2 / China (Beijing), and create an ordinary Model Studio key on the Beijing API Key page. Earlier tutorials in this series walk through that signup in more detail. This page assumes you already have a Beijing key and a workspace ID, and that you have flipped 免费额度用完即停 on for each of qwen3.8-flash, qwen3.8-max, and qwen4-max on the Beijing free quota page.
Use a regular Model Studio key, not a Token Plan or Coding Plan dedicated key. Official free-quota docs say those dedicated keys do not consume the new-user grant. Prefer the documented workspace host, and keep the key and the host in the same region:
export DASHSCOPE_API_KEY="sk-..."
export QWEN_BASE_URL="https://<WorkspaceId>.cn-beijing.maas.aliyuncs.com/compatible-mode/v1"
Official compatibility docs still describe the older host https://dashscope.aliyuncs.com/compatible-mode/v1 as working. Either form stops at /compatible-mode/v1; the OpenAI SDK appends /chat/completions.
cd examples
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python compare_three_quotas.py
Planned command. Not run for this draft. If the two environment variables are missing, the script prints the judging plan and exits 2. It does not invent scores. If you already use Codex or Grok Bot, you can paste the script there after the variables are set. You do not need a second coding-assistant login to follow these steps.
The call itself is one Chat Completions request per model:
completion = client.chat.completions.create(
model=model, # qwen3.8-flash, then qwen3.8-max, then qwen4-max
messages=[
{"role": "system", "content": SYSTEM},
{"role": "user", "content": USER_TEMPLATE.format(changelog=CHANGELOG)},
],
extra_body={"enable_thinking": False},
)
The checker accepts a bare object or one wrapped in markdown fences, then compares paths and types with examples/expected.json. A later verified revision should fill this table from examples/results/summary.json. Until then, the cells stay empty on purpose.
| Model | Checker score | Clean pass | Invented breaking changes | prompt_tokens | completion_tokens | total_tokens | request id |
|---|---|---|---|---|---|---|---|
qwen3.8-flash |
not yet run | not yet run | not yet run | not yet run | not yet run | not yet run | not yet run |
qwen3.8-max |
not yet run | not yet run | not yet run | not yet run | not yet run | not yet run | not yet run |
qwen4-max |
not yet run | not yet run | not yet run | not yet run | not yet run | not yet run | not yet run |
Do not read the empty cells as a tie, and do not borrow a winner from a benchmark leaderboard. Third-party comparison pages disagree with each other, and some of them misstate official limits — including context size and whether qwen3.8-max accepts video. The cards in the next section are taken from Model Studio model pages and the China-site price table, not from those scoreboards.
Do the three Beijing grants share a token pool?
No. The new-user quota rules say each model has its own grant, usually 1,000,000 tokens, and that grants cannot be merged or moved. A snapshot with a date suffix is a different model from the unsuffixed alias. When one grant is exhausted, the service does not fail over to another ID; you change the model parameter yourself if you want to spend a different pool.
That is why “which free quota to use” is not the same question as “which model is free.” All three listed IDs have a Beijing grant on the China-site pricing table as of 2026-11-17. You can, in principle, spend all three. The snapshot qwen3.8-max-0902 (also documented as qwen3.8-max-2026-09-02) is the trap in this family: it is billed at the same Beijing list price as qwen3.8-max, and the price table marks its free quota as 无 — none.
Input tokens and output tokens, including thinking tokens, draw from the same remainder for that model. The console keeps one running total. Only live inference is in scope; batch jobs, fine-tuning, and deployment are billed separately, if they are available at all. The documented remainder is shared by the primary Alibaba Cloud account and every RAM user under it. Official rules also say a new account under the same real-name identity does not get another new-user grant.
On the China site, these grants are offered only in 华北2 (Beijing). Other China-site regions show list prices and no grant column. The international-site English pricing pages invert that geography: they offer a 1,000,000-token grant in Singapore (International scope) and say other regions have none. If you signed up on aliyun.com from mainland China, stay in Beijing. The Singapore grant on the English site is not attached to a China-site Beijing key.
Which of these grants may already have expired?
Look at the console expiry. Do not assume that a price table which still lists a grant means your particular account still has one.
Official rules start the 90-day clock at the later of Model Studio activation, the model’s release, or application approval. The model listing page records:
| Model ID | Official listing date | 90 days after that date |
|---|---|---|
qwen3.8-max |
2026-08-02 | 2026-10-31 |
qwen3.8-flash |
2026-08-26 | 2026-11-24 |
qwen4-max |
2026-11-16 | 2027-02-14 |
If the account was already open on the listing date, the clock can start on that date. As of 17 November 2026, that reading means a qwen3.8-max grant on such an account may already have expired, a qwen3.8-flash grant may have about a week left, and a qwen4-max grant is still early in its window. If you activated Model Studio after a model was listed, the later date is the activation date, and the console may show a later expiry than the table above.
This page has not opened a customer console, so it cannot tell you which of those two cases you are in. It also cannot confirm whether a brand-new China-site activation on 17 November 2026 still receives all three grants. The price table still lists them. The authority for an individual account is the 免费额度 page.
What do the official cards actually differ on?
For a same-task text extraction, the three Beijing cards are closer than the marketing names suggest. All three official model pages list a 1,000,000-token context window, a 991,808-token maximum input (983,616 in thinking mode), text, image, and video input, text output, function calling, structured output, web search, prefix completion, and context cache. Fine-tuning is listed as unsupported on all three. Thinking is hybrid and on by default. Beijing rate limits are described as dynamic TPM tiers tied to monthly Model Studio spend.
The differences that are actually on the official cards, read 2026-11-17:
qwen3.8-flash |
qwen3.8-max |
qwen4-max |
|
|---|---|---|---|
| Official listing | 2026-08-26 | 2026-08-02 | 2026-11-16 |
| Beijing free quota | 1,000,000 tokens, 90 days | 1,000,000 tokens, 90 days | 1,000,000 tokens, 90 days |
| Beijing list price, per 1M tokens | CNY 0.8 in / 2.7 out | CNY 12 in / 36 out | CNY 10 in / 30 out |
| Cache, Beijing, per 1M tokens | Hit CNY 0.1; explicit create CNY 1.25 | Implicit hit CNY 1.5; explicit create CNY 15; explicit hit CNY 1 | 5-minute create CNY 12.5; 1-hour create CNY 20; explicit/implicit hit CNY 0.5 |
| Max answer length | 131,072, thinking or not | 131,072, thinking or not | 131,072 with thinking off; 262,144 for thinking plus the answer |
| Max chain-of-thought length | 262,144 | 262,144 | 262,144 |
| Batch inference in Beijing | Listed as supported on the model page. The price row does not carry a “Batch half-price” badge, and the model page does not print batch unit prices. | Supported. Batch File is listed at CNY 6 / 18; Batch Chat stays at CNY 12 / 36. | Not supported |
| Official first-choice status | Cost-first pick on the text-generation guide | Filed under older models | Quality-first pick on the text-generation guide |
| Other documented extras | — | Snapshot qwen3.8-max-0902 exists, with no free quota |
reasoning_effort of low, high, or xhigh (default high); default thinking_budget 131,072; chain-of-thought summary and encryption; parallel tool calls |
Singapore prices on the same China-site table, for readers who later leave Beijing, have no free quota: qwen3.8-flash at CNY 1.094 / 3.427, qwen3.8-max at CNY 14.988 / 44.965, and qwen4-max at CNY 10.938 / 32.813, per million tokens.
Once a Beijing grant is gone, and only then, a short arithmetic example (not a measured bill) helps fix the scale. For 20,000 input tokens and 2,000 output tokens, with thinking off: qwen3.8-flash would be CNY 0.0214, qwen4-max CNY 0.26, and qwen3.8-max CNY 0.312. Inside the free grant, that same request simply deducts 22,000 tokens from that model’s remainder.
How should you choose before a live run exists?
Spend the grant that matches the job, after you have checked that the grant is still there. Do not spend the flagship pools on work the Flash grant can finish, and do not assume qwen3.8-max is still the default Max model just because the name says Max.
A practical order, based on the official cards rather than on unrun outputs:
- Confirm each model’s remaining tokens and expiry on the Beijing free-quota page. If
qwen3.8-maxalready shows as expired, drop it from the plan without treating that as a quality verdict. - Send cheap, checkable jobs — extraction, classification, short drafts, first-pass summaries — to
qwen3.8-flashfirst. Official guidance already says to confirm that the result is good enough before you standardize on it. - Reserve
qwen4-maxfor the cases you would have called a flagship for anyway: harder reasoning, longer agent traces, or answers that may need the larger thinking-mode budget. After the grant, it is also the less expensive of the two Max IDs in Beijing. - Spend
qwen3.8-maxwhen its grant is still live and you want a second flagship pool, or when you are already mid-evaluation on that generation. After the grant, the list price is the highest of the three, and official positioning has moved on. - Do not point this comparison at
qwen3.8-max-0902if the goal is to stay on a free grant. Use that snapshot only when you need a pinned 3.8-Max revision and you have accepted list price.
Until the results table is filled, treat Flash as the default for volume, Qwen4-Max as the default for difficulty, and 3.8-Max as a residual pool with a clock you have to inspect. That is a reading of the official cards and the quota rules. It is not a claim about which model will score higher on the Quarrypond changelog.
When is one of these the wrong grant to spend?
- The snapshot ID.
qwen3.8-max-0902/qwen3.8-max-2026-09-02has no new-user grant on the China-site price table. It is the wrong string if you came here for a free quota. - A Token Plan-only preview ID. This comparison uses
qwen4-max, the generally available Model Studio ID. It does not useqwen4-max-preview, which official docs restrict to Token Plan keys. - An expired bar that you ignored. A model that still appears in the catalog can have a dead new-user grant. The console expiry is the check; the listing date is only a way to know that the check is worth doing.
- A job that needs batch inference on the free grant. The grant does not cover batch.
qwen4-maxdoes not list batch support at all. If you need Batch File pricing, that is a paidqwen3.8-maxpath, not a free-quota comparison. - A job that needs more than 131,072 tokens of answer with thinking off. All three models share that non-thinking cap. The extra 262,144-token thinking-plus-answer budget is documented only for
qwen4-max, and only with thinking on. - An international-site account. Singapore grants and USD prices are a different signup. This page stays on aliyun.com and Beijing so the steps can be finished from mainland China.
- Production traffic you cannot afford to stop. Official docs warn that 免费额度用完即停 will refuse further calls once the grant is exhausted. That is the correct setting for a zero-spend comparison. It is the wrong setting if a live path must keep going and you have accepted pay-as-you-go.
What a verified revision has to add
A later revision should replace the unverified banner with the run date, the openai package version, the three request IDs, the usage objects as returned, and the checker scores from examples/results/summary.json. It should also record whether each grant still appeared in the console on that day, whether thinking-off responses stayed inside message.content as documented, and whether a first-time China-site activation on or after 17 November 2026 still received all three 1,000,000-token grants.
Until that run exists, treat the quality ranking as unknown and the official cards as the part of this page you can cite.
Sources
Official Model Studio documentation, retrieved 2026-11-17 (Asia/Shanghai):
- Model pricing (China site)
- Free quota for new users
- Newly released models
- qwen3.8-flash
- qwen3.8-max
- qwen4-max
- Text generation and visual understanding
- Deep thinking
- OpenAI compatibility
International-site pricing and quota pages were opened only to confirm the site split, not as the path this tutorial uses:
Third-party comparison pages read for search context, not as a source of limits or winners:
- Qwen3.8 Flash vs Qwen4-Max (llm-stats)
- Qwen3.8 Flash vs Qwen3.8 Max (llm-stats)
- Qwen 4 vs Qwen 3.8 Max (OrcaRouter)
- Qwen3.8 Flash vs Qwen3.8 Max (SynthorAI)
Title tag: Qwen3.8-Flash vs Qwen3.8-Max vs Qwen4-Max Beijing Free Quotas
Meta description: Which Beijing free quota to spend first on qwen3.8-flash, qwen3.8-max, or qwen4-max. Official prices and 90-day clocks; the same-task run is not yet verified.
Slug: qwen3.8-flash-vs-qwen3.8-max-vs-qwen4-max-free-quota
Evidence it actually ran
This is the shape of the stamp. It has not been earned, and the example date that follows is a placeholder, not a run: Ran on YYYY-MM-DD (Asia/Shanghai) with the pinned versions.
- Command output
- None recorded. A verified revision will quote the output from the run, not a sample written in advance.
- Tests, CI, or Inspector
- Not run for this draft.
- Repository or config
- No link recorded.
Failure modes
Anticipated only. None of these has been checked against a run of this tutorial.
Unchecked
A snapshot ID is sent instead of the granted model ID
qwen3.8-max-0902 / qwen3.8-max-2026-09-02 has no new-user grant on the China-site price table. It is the wrong string if you came here for a free quota.
Unchecked
A Token Plan-only preview ID is used
This comparison uses qwen4-max, the generally available Model Studio ID. It does not use qwen4-max-preview, which official docs restrict to Token Plan keys.
Unchecked
An expired grant bar is ignored
A model that still appears in the catalog can have a dead new-user grant. The console expiry is the check; the listing date is only a way to know that the check is worth doing.
Unchecked
The job is treated as a free batch comparison
The grant does not cover batch. qwen4-max does not list batch support at all. If you need Batch File pricing, that is a paid qwen3.8-max path, not a free-quota comparison.
Unchecked
The three grants are assumed to share one token pool
Each eligible model ID gets its own new-user grant. Exhausting qwen4-max does not automatically start spending qwen3.8-flash.
Unchecked
免费额度用完即停 is left off on a comparison you wanted to keep free
Official docs warn that the switch will refuse further calls once the grant is exhausted. That is the correct setting for a zero-spend comparison. Leave it off, and an exhausted grant flips to list price.
Related
Other tutorials in Integrate.
Connect a Local MCP Server to Cursor and Use It to Finish a Real Repo Task
Start a local MCP server, connect it in Cursor, and complete a repo task that requires that tool.
Not yet verified
1 min read
Add a Second Frontier Model to Claude Code with a Documented Fallback Path
Configure a second model and complete the same task via the fallback after a simulated failure.
Not yet verified
1 min read
Build a TypeScript MCP Server for a Real API, Test It with Inspector, and Call It from Claude Code
Implement the server, pass Inspector, then call it from Claude Code to finish a real API task.
Not yet verified
1 min read
Call Qwen4-Max Through the OpenAI-Compatible API and Finish a Tool-Calling Task
Call qwen4-max from the OpenAI SDK on Alibaba Cloud Model Studio in Beijing by changing only api_key, base_url, and model, stay on that model's official free quota with 免费额度用完即停 / Free Quota Only turned on, then finish a tool-calling job: a checked dependency report for an npm project.
Not yet verified
15 min read