Tag: Alibaba Cloud

  • Alibaba Launches Qwen4-Max on the QwenCloud API (2026)

    Alibaba Launches Qwen4-Max on the QwenCloud API (2026)

    Takeaway: On November 16, 2026, Alibaba’s Qwen team made Qwen4-Max generally available on the QwenCloud API. The model ID is qwen4-max. QwenCloud’s documentation calls it the first model in the Qwen4 family: a flagship multimodal system for text generation, image and video understanding, reasoning, coding, and agent tasks, with a 1,000,000-token context window. Official list prices on the Singapore (International) table are $1.50 per million input tokens and $4.50 per million output tokens. On the Global table, including China North 2 (Beijing), the same model is listed at $1.413 and $4.24. A Token Plan-only preview, qwen4-max-preview, had already appeared on November 11. This piece sticks to QwenCloud and Alibaba Cloud documentation for specs, prices, and limits.

    The launch lands in a busy week for flagship model access. Google only just opened Gemini 4 Carbon to paying developers, and pricing remains a practical buying question for teams comparing lower-cost frontier alternatives. Qwen4-Max is a different product: a QwenCloud API model with OpenAI-compatible, Anthropic Messages, and DashScope endpoints, thinking enabled by default, and an explicit cache designed for this generation.

    What Qwen says Qwen4-Max is for

    QwenCloud’s latest-model page describes Qwen4-Max as combining reasoning, native vision, and long-running task execution. It lists four job types. For coding and engineering, the documentation says the launch blog demonstrates building an Unreal Engine game, turning production-line layouts into 3D models, and improving machining plans through simulation feedback. For professional work, the examples are financial research with linked Excel workbooks and editable presentations, and legal due diligence with document classification, cross-source checks, and traceable citations. For long-running agents, it points to the Qwen Live Delivery plugin, which it says can deliver a usable first version and update it as work progresses. For scientific exploration, the same page says the blog describes a cooling-fluid molecule designed and synthesized through an integrated experimental platform, and drug candidates optimized through computational evaluation for further laboratory validation.

    Those examples are Qwen’s own illustrations, not independent third-party evaluations. The documentation itself tells readers to see the launch blog for the examples and full evaluation results. Alibaba Cloud’s Model Studio release notes, updated November 16, 2026, add that the model supports web search, tool calling, and structured outputs, and that it is compatible with the OpenAI and Anthropic APIs. The QwenCloud model changelog records the same date and describes Qwen4-Max as a flagship multimodal model for text, image and video understanding, reasoning, coding, and agent tasks, with tool calling, structured outputs, explicit caching, and chain-of-thought summaries.

    Specs, defaults, and hard limits

    The official specifications table is more useful than a marketing one-liner. Context length is 1,000,000 tokens. In non-thinking mode, maximum input is 991,808 tokens and maximum output is 131,072 tokens. In thinking mode, maximum input is 983,616 tokens, and thinking plus the final answer can use up to 262,144 tokens. The default thinking budget is 131,072 tokens. Thinking is on by default, with reasoning_effort=high.

    Supported capabilities on the same page include text generation, thinking, function calling, structured output, image understanding, video understanding, explicit cache, CoT Summary, web search, PDF understanding, a code interpreter, web scraping, text-to-image search, image-to-image search, and MCP. Streaming, parallel tool calls, prefix completion, and implicit caching are also supported. The code interpreter requires thinking to be enabled. Batch inference is listed as not supported. Fine-tuning and model deployment are also listed as not supported.

    Default sampling changes with the mode. In thinking mode, temperature defaults to 1.0, top_p to 0.95, and presence_penalty to 0.0. In non-thinking mode, those defaults are 0.7, 0.8, and 1.5. Top_k is 20 and repetition_penalty is 1.0 in both modes. In Chat Completions, enable_thinking=false turns thinking off. max_completion_tokens caps thinking and the final answer together. preserve_thinking is enabled by default, and QwenCloud says you should keep assistant messages and their reasoning in multi-turn tool calls.

    For images and video, image detail defaults to high. An explicit max_pixels takes priority, including when vl_high_resolution_images=true. Video total_pixels has a maximum of 268,435,456. The documentation also describes a CoT Summary option that returns a readable summary of the model’s reasoning plus encrypted full reasoning, rather than the full chain of thought in plain text. The summary is for display; the encrypted content is what restores reasoning history in later turns. QwenCloud warns that a summary cannot replace that encrypted history, and that summaries or encrypted content can be absent.

    Pricing, cache, and the Singapore free quota

    Prices depend on the region and service deployment scope. On the official model pricing table, qwen4-max uses a single input tier of up to 1 million tokens per request, and the same input and output rates apply in thinking and non-thinking modes. Output pricing covers the chain of thought plus the answer. Singapore (International) is listed at $1.50 per million input tokens and $4.50 per million output tokens, with a free quota of 1 million tokens. Global prices, which QwenCloud says are shared with China North 2 (Beijing), are $1.413 and $4.24. On that same Global table, qwen3.8-max is listed at $1.65 and $4.951, so the new flagship’s published Global rates sit below its predecessor’s on the official sheet. That is a list-price comparison, not a promise that a given task will cost less.

    Cache has its own rates. For Singapore, QwenCloud lists cache writes at $1.875 per million tokens for a 5-minute time-to-live and $3 for a 1-hour TTL, with explicit or implicit cache reads at $0.075. For Global, the same three figures are $1.767, $2.827, and $0.071. The latest-model documentation says the new Qwen4 explicit cache can store shared prefixes such as system prompts, tool definitions, and reference documents. TTL options are 5m (the default) and 1h. Implicit cache does not guarantee hits. Chat Completions, Responses, and Anthropic Messages support the new explicit cache; DashScope does not.

    QwenCloud’s pricing guide says new users in the Singapore region get a free quota and that other regions have none. The latest-model page adds that pay-as-you-go calls in Singapore offer 1 million free tokens, separate from Token Plan credits, and points to the free-quota documentation for eligibility and validity. Failed API calls are not charged and do not consume free quota. Always confirm the live table before you budget a workload, because regional rates and quota rules can change.

    How you call it, and where it runs

    The documented model ID is qwen4-max. A Singapore quick start uses the OpenAI SDK against https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1, with a DASHSCOPE_API_KEY. QwenCloud also documents Responses, Anthropic Messages, and DashScope. It notes that Qwen4-Max uses the DashScope multimodal endpoint even for text-only input. Available regions listed on the latest-model page are Singapore (International); China North 2 (Beijing, China); and Frankfurt, Hong Kong, Tokyo, and US (Virginia) under the Global scope.

    Reasoning effort maps slightly differently across protocols. OpenAI Chat Completions and DashScope use reasoning_effort with low, high, or xhigh (default high). OpenAI Responses uses reasoning.effort with the same values. Anthropic Messages uses output_config.effort. Compatibility aliases include minimal to low, medium to high, and max to xhigh. The documentation also says you can use Qwen4-Max through Token Plan Personal, Token Plan Team, or the pay-as-you-go API, and that Token Plan usage consumes credits through the corresponding API key and endpoint. It lists configuration notes for Claude Code, Codex, and OpenCode.

    Preview last week, general API access today

    The changelog entry for November 11, 2026 introduced qwen4-max-preview. Alibaba Cloud’s Model Studio notes from that date said the preview was available through Token Plan only. The November 16 entry is a different model ID, qwen4-max, with published pay-as-you-go prices and a public latest-model guide. QwenCloud still says the separate preview model has its own preview benefits, so teams already mid-preview should check which ID and credential path they are on before switching traffic.

    One caution is worth repeating because product pages can drift from the developer guide. The official latest-model documentation is explicit that batch inference is not supported and that fine-tuning and model deployment are not supported. If a marketplace checklist shows those items as features, treat the developer specification as the rule until QwenCloud changes it in the same docs.

    What Alibaba’s official launch post adds

    QwenCloud’s developer documentation is the source for the prices, context limits, and API details in this article. A separate official Alibaba Cloud Community post dated November 16, 2026, adds architecture and evaluation claims that the QwenCloud docs summarize only at a high level. That post describes Qwen4-Max as a sparse mixture-of-experts model with 2.4 trillion total parameters that activates 71 billion per token — 25% fewer activated parameters than Qwen3.8-Max at the same 2.4 trillion scale, according to Alibaba. It also says Qwen4-Max ranked 6th on Code Arena and, on Alibaba’s proprietary Work360 benchmark, outperformed Qwen3.8-Max overall while using 35% fewer tokens at High effort reasoning mode and reducing average cost per task by around 60%. Those figures are Alibaba’s, not independent third-party results. The qwen.ai launch blog cited by QwenCloud for examples and full evaluation results is a JavaScript page; this update relies on the Alibaba Cloud Community post and QwenCloud documentation rather than an incomplete scrape of that blog.

    Availability at a glance

    ItemWhat official Qwen / Alibaba Cloud docs say (November 16, 2026)
    What launchedQwen4-Max, the first model in the Qwen4 family, generally available on the QwenCloud API
    Model IDqwen4-max
    Official datesNovember 16, 2026 general availability; November 11, 2026 Token Plan preview as qwen4-max-preview
    Context window1,000,000 tokens
    Non-thinking limitsMaximum input 991,808 tokens; maximum output 131,072 tokens
    Thinking limitsMaximum input 983,616 tokens; thinking plus final answer up to 262,144 tokens; default thinking budget 131,072 tokens
    Thinking defaultEnabled, with reasoning_effort=high
    Singapore list price$1.50 / $4.50 per million input / output tokens; cache write $1.875 (5m) and $3 (1h); cache read $0.075
    Global / Beijing list price$1.413 / $4.24 per million input / output tokens; cache write $1.767 (5m) and $2.827 (1h); cache read $0.071
    Singapore free quota1 million tokens on pay-as-you-go, separate from Token Plan credits; other regions have no free quota
    APIsOpenAI-compatible Chat Completions and Responses, Anthropic Messages, and DashScope
    Singapore endpoint examplehttps://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1
    Regions named in the docsSingapore (International); China North 2 (Beijing); Frankfurt, Hong Kong, Tokyo, and US (Virginia) (Global)
    Not supported, per the latest-model pageBatch inference; fine-tuning; model deployment
    New explicit cacheSupported on Chat Completions, Responses, and Anthropic Messages; not supported on DashScope

    FAQ

    Is Qwen4-Max the entire Qwen4 family?

    No. QwenCloud calls it the first model in the Qwen4 family. The publicly documented API model launching today is qwen4-max. This article does not treat unreleased siblings or open-weight plans as facts unless Qwen’s own documentation states them.

    How much does Qwen4-Max cost?

    Official Singapore (International) list prices are $1.50 per million input tokens and $4.50 per million output tokens, including thinking and non-thinking modes, for requests up to 1 million input tokens. Global prices, shared with Beijing, are $1.413 and $4.24. Cache writes and reads are billed at the published cache rates. Output pricing includes the chain of thought plus the answer. Confirm the live pricing table for your region before you ship.

    Is there a free way to try it?

    QwenCloud says pay-as-you-go calls in Singapore include 1 million free tokens, separate from Token Plan credits, and that other regions have no free quota. Eligibility and validity are spelled out in the free-quota documentation. The November 11 preview was Token Plan only and used a different model ID.

    Does thinking stay on unless I turn it off?

    Yes. Thinking is enabled by default, with a default effort of high and a default thinking budget of 131,072 tokens. In Chat Completions, set enable_thinking=false to turn it off. The code interpreter requires thinking to be enabled. Remember that thinking tokens are part of the billed output.

    Can I fine-tune Qwen4-Max or run batch jobs?

    Not according to the official latest-model specification. Batch inference is listed as not supported, and fine-tuning and model deployment are listed as not supported. Use the pay-as-you-go or Token Plan APIs for online inference, and watch the changelog if that policy changes.

    Which APIs and tools are compatible?

    QwenCloud documents OpenAI-compatible Chat Completions and Responses, Anthropic Messages, and DashScope, plus notes for Claude Code, Codex, and OpenCode. Function calling, structured outputs, web search, MCP, and several built-in tools are supported, with protocol-specific limits. The new Qwen4 explicit cache is not available on DashScope.

    How is this different from last week’s preview?

    On November 11, Alibaba Cloud documented qwen4-max-preview as a Token Plan-only preview. November 16 is the general qwen4-max launch, with published pay-as-you-go prices, a latest-model developer guide, and a changelog entry. QwenCloud still refers to separate preview benefits on the preview ID, so check your model ID and key path before you cut over.

    Where should I look for official evaluation numbers?

    Start with QwenCloud’s latest-model page, the launch blog it cites, and Alibaba’s official November 16 Cloud Community post. The parameter counts, Code Arena placement, and Work360 comparisons in this article are attributed to that Alibaba post. API prices and limits still come from QwenCloud’s own documentation. Figures that appear only in secondary coverage are omitted unless we can attribute them to those official documents.

    As of this writing on November 16, 2026, Qwen4-Max is live in QwenCloud’s changelog, Model Studio release notes, pricing tables, and latest-model guide. If you are choosing a flagship API model this week, the practical next step is to read those pages for your region, decide whether you want thinking on by default, and price a real workload against the published token and cache rates rather than a headline number.