AdvertiseFee Deal Desk

Macro

Alibaba Launches Qwen4-Max Flagship With 2.4 Trillion Parameters

Alibaba launched its Qwen4-Max flagship on November 16, 2026, a 2.4-trillion-parameter model priced from $1.50 per million input tokens in Singapore.

Be a creator
November 16, 2026, 09:42 AM UTC4 min read
AI SummaryAI
  • The mixture-of-experts model holds 2.4 trillion total parameters and activates about 71 billion per token.
  • Qwen4-Max accepts text, image and video input and supports up to 1 million tokens of context.
  • Singapore pay-as-you-go pricing lists $1.50 per million input tokens and $4.50 per million output tokens.
  • Cache reads cost $0.075 per million tokens, and eligible users get 1 million free tokens for 90 days.
accszone.com

What Alibaba's Announcement Specifies

Alibaba released Qwen4-Max on Monday, November 16, presenting it as the flagship model of the Tongyi Qwen family, aimed at programming, workplace document processing and multi-step agent tasks. The launch announcement describes a mixture-of-experts architecture with 2.4 trillion total parameters, of which roughly 71 billion activate on any single token. That design does not put the full parameter count to work on every answer, and Alibaba's own materials caution against judging output quality on parameter totals alone.

The model accepts text, images and video, and returns text. Official documentation lists a context window of up to 1 million tokens, a ceiling on how much material one session can hold rather than a promise of output length. It also lists reasoning with adjustable thinking intensity, tool calling including parallel calls and web search, structured output via JSON Object and JSON Schema, and interfaces compatible with OpenAI and Anthropic alongside Alibaba Cloud's DashScope API. Output ceilings differ by mode: 131,072 tokens in non-thinking mode, against 262,144 tokens covering both reasoning and the final reply in thinking mode, figures the documents say should not be conflated.

The announcement frames the model as one member of a wider series. Qwen is Alibaba's model family, known in Chinese as Tongyi Qwen, and Qwen4-Max is its flagship entry rather than a name covering every Qwen product. The materials separate three layers: the model itself, the chat and work products that run it, and the API developers call. Tool support, the documents add, does not mean a model name typed into a prompt automatically gains access to a computer, an account or company data; what a deployment can do depends on the product, the tool configuration and the permissions granted. Alibaba positions the release for coding, office work and research, advising buyers to test against their own data, costs and error tolerance rather than leaderboard rankings. OpenAI compatibility, it notes, is a protocol match, not use of OpenAI models or OpenAI billing.

Singapore Pricing and the Free Quota

Access runs through three entry points named in the announcement: QwenStudio for direct conversation, QwenWork for task handling, and QwenCloud, which sells the API and subscription plans. The production model ID is qwen4-max. Under Alibaba Cloud Model Studio, developers must match the API key, workspace and service region, and settings tied to a Token Plan do not carry over to the standard pay-as-you-go API.

Singapore pay-as-you-go pricing for the International deployment scope, listed on November 16, runs $1.50 per million input tokens and $4.50 per million output tokens. Cache reads cost $0.075 per million tokens, while explicit cache writes are priced at $1.875 for a five-minute window and $3.00 for one hour. The cache-read figure applies only to requests that hit a cache; uncached input bills at the full $1.50, and creating an explicit cache carries its own write fee. On the listed prices, 100,000 uncached input tokens plus 10,000 billed output tokens cost about $0.195.

A free quota of 1 million tokens covers eligible pay-as-you-go calls in the Singapore International scope. It stays valid for 90 days from the latest of three dates: service activation, the model release, or approval of an application. Token Plan keys do not draw on it. Once the quota is spent or expires, continued use can roll into paid billing; the Free Quota Only switch ships off by default, and changes to it do not take effect immediately.

The preview version, qwen4-max-preview, entered the changelog on November 11 and is available only through the Token Plan. Its preview window runs from November 11, 10:00 to November 25, 24:00, UTC+8, after which the documents say it retires. The official version joined the formal model updates on November 16. Alibaba also notes that smaller Qwen4 open-weight models are planned, which does not make Qwen4-Max weights available for download.

Cache Pricing Shapes the Bill

The pricing spread is the number that decides real-world costs for heavy users. A workload dominated by repeated, cacheable context bills near $0.075 per million input tokens, while one that misses cache pays twenty times that, and the announcement's own example shows how quickly the two paths diverge. Output at $4.50 per million tokens runs at triple the input rate, so mode choice matters too: thinking mode doubles the combined output ceiling to 262,144 tokens against the 131,072 available without it. Teams still on the preview have until November 25 to retest parameters and output limits against the production ID, and the planned smaller Qwen4 open-weight models give cost-sensitive workloads an alternative to the flagship.

COINOTAG's editorial and research desk.

AI-Assisted

AI-generated, AI-reviewed, under COINOTAG editorial oversight.

Price-Impacting News

Live
All breaking news
Breaking News