what is qwen4-max? alibaba's new flagship model, explained
qwen4-max is alibaba's new flagship model in the qwen family, released on november 16, 2026: a mixture of experts (moe) model with 2.4 trillion total parameters, native vision and a context window of one million tokens. alibaba says it is available now on qwencloud via the token plan and api, and in alibaba platforms including qwenwork and qwenstudio.
alibaba calls it the most capable model in the qwen family to date, and its announcement says it excels at complex tasks across work, creativity and research.
download the ReMynd desktop app to search text you saw in a focused window, by app and date.
what did alibaba announce?
alibaba announced qwen4-max in a press release from hangzhou, china. the release makes these claims:
- it ranked 6th on code arena, which alibaba calls a real-world coding leaderboard.
- on work360, a proprietary alibaba benchmark of challenging industry tasks across 721 specialized fields, it outperforms qwen3.8-max overall.
- on that benchmark it uses 35% fewer tokens at high effort reasoning mode and cuts the average cost per task by around 60%.
- it posted strong scores on agents' last exam, vals index, jobbench and chartography, according to alibaba, which gives no scores for them in the release.
these are alibaba's own results. a later alibaba cloud community blog post, dated november 17, gives the more precise figure for that work360 cost cut: 58%. the release also says open-weight, smaller-scale models from the qwen4 series will soon be released to developers worldwide.
how big is qwen4-max?
qwen4-max is a sparse moe model, so only part of its parameters work on each token. alibaba says the new qwen4 architecture upgrades that design across attention, residual connections and embeddings.
| spec | what alibaba lists |
|---|---|
| activated per token | 71 billion, 25% fewer than qwen3.8-max at the same scale |
| maximum input | 991,808 tokens (983,616 in thinking mode) |
| maximum output | 131,072 tokens (262,144 in thinking mode) |
| input / output | text, image and video in; text out |
| reasoning effort | low, high or xhigh; high is the default |
the activated row comes from alibaba's announcement. the token limits, modalities and reasoning effort rows come from the alibaba cloud model studio qwen4-max page, last updated november 21, 2026. it lists deep thinking, parallel tool calls, web search, function calling, structured output and explicit and implicit cache as supported.
how much does qwen4-max cost?
qwen4-max is billed per token through the api, and the price depends on the region. qwencloud's model page lists these api list prices:
| price per 1m tokens | qwencloud |
|---|---|
| input | $1.5 |
| output | $4.5 |
| input, implicit cache | $0.075 |
| 5-minute explicit cache creation | $1.875 |
| 1-hour explicit cache creation | $3 |
| explicit cache read | the same rate as an implicit cache hit |
model studio lists the singapore (international) rate at the qwencloud figures above, with a free quota of 1 million tokens valid within 90 days of activation, model release or application approval, whichever is later.
model studio lists $1.413 input and $4.24 output in china (beijing), hong kong (china), germany (frankfurt), united states (virginia) and japan (tokyo). united states (virginia) also lists a second, virginia deployment scope at the qwencloud rates above. model studio lists batch inference as supported in china (beijing), at $0.707 input and $2.12 output for batch file calls, and as not supported in its other region tables.
for scale, the model studio pricing page lists qwen3.8-max in singapore at $2 input and $6 output per 1m tokens.
how do you call qwen4-max?
you call qwen4-max by its model id, qwen4-max, through one of four call methods. alibaba's model studio page lists them:
- openai chat completions.
- openai responses.
- anthropic messages.
- dashscope.
qwencloud's model page says qwen4-max is compatible with the openai and anthropic apis, and lists a limit of 5000000 tokens per minute there. model studio adds a limit of 30,000 requests per minute in its global deployment regions, and lists lower limits for the virginia deployment scope.
where does ReMynd fit?
ReMynd records your screen on a mac, runs ocr locally with apple vision, and keeps a searchable index on your computer. text comes from the focused window only, and video covers windows you have not excluded.
in our view, even a very long context window holds only what gets sent to the model, so it does not know what you read on your screen last week unless you send it. ReMynd keeps that history searchable: press control space, search in plain language, and jump to the exact moment, provided ReMynd was running then and the window was not excluded. on a mac it needs macos 14 or later on an apple silicon mac.
ReMynd also runs on windows 11 (x64 or arm64). for another new model, see what is aion instruct.
limits of this guide
- it follows alibaba's launch press release, alibaba cloud's community blog post, qwencloud's model page and alibaba cloud model studio's docs, as of november 2026. we did not run the model.
- the benchmark claims are alibaba's own, and the release gives a rank and percentages, not full score tables.
- prices are alibaba's list prices by region; alibaba's pricing page says it shows only the original prices, and promotional offers appear in the model studio console.
download the ReMynd desktop app and search what you saw in a focused window, by app and date.