Skip to main content
株式会社オブライト
株式会社オブライト

Qwen4-Max: API Pricing, 1M Context & Closed Weights

Published 2026/11/169 min read

Qwen4-Max is Alibaba's top API model (GA Nov 16): text, image, video input, 1M context, $1.50 in and $4.50 out per 1M tokens. Weights closed; no local run.

Qwen4-Max is the top-tier large language model that Alibaba's Qwen team made generally available on November 16, 2026. It accepts text, image and video input, produces text output, and has a 1-million-token context window. The official QwenCloud model page lists $1.50 per million input tokens and $4.50 per million output tokens. In short: it is an API-only, closed-weight model, so it cannot be run locally.

A limited preview (qwen4-max-preview) for Token Plan users came first on November 11; general availability opens normal API access. Scale figures such as "2.4 trillion total / 71B active parameters" could not be confirmed on an official page at the time of writing, so we treat them as press-reported. This article separates what we verified on official pages from what is reported.

The practical takeaway: because the API is OpenAI-compatible, you can try it by swapping the base_url and model name in existing code. Watch out for data residency, confusion with the preview model ID, and inconsistencies between official documents on some limits.

What Qwen4-Max Can Do

According to the official model page, Qwen4-Max accepts multimodal input including images and video, and supports function calling, structured output, context caching and web search. Output is text only; it does not generate audio or images. Typical uses are reading long documents or many files in a single request and agent-style work that calls tools.

ItemDetailSource
ReleaseNov 16, 2026 (GA); limited preview Nov 11Press (CASRAI)
Model IDqwen4-max (qwen/qwen4-max on OpenRouter)Official model page / OpenRouter
InputText, image, videoOfficial model page
OutputTextOfficial model page
Context1M tokensOfficial model page
Max input~991K tokens (~983K with thinking)Official model page
Max reasoning~262K tokensOfficial model page
Rate limit5,000,000 tokens/min (TPM)Official model page
ArchitectureMoE, 2.4T total / 71B active parametersPress-reported (CASRAI, attributed to Alibaba)
WeightsNot released (no download)Press-reported (CASRAI)

Rows marked "Official model page" come from the QwenCloud model page (https://www.qwencloud.com/models/qwen4-max). The parameter counts and closed-weights status are press-reported because we could not directly read the official blog post.

Note that sources disagree on maximum output: the model page shows 393K, while the developer guide quoted by press says 131,072 (non-thinking). Support for batch inference and fine-tuning is also described differently. Confirm the latest values in your own console before relying on them.

API Pricing (Official Model Page Values)

CategoryPrice per 1M tokens
Input$1.50
Output$4.50
Implicit cache input$0.075
Explicit cache creation (5 min)$1.875
Explicit cache creation (1 hour)$3.00
Explicit cache read$0.075

These are the values on the QwenCloud (Singapore endpoint) model page. OpenRouter's page likewise lists $1.50 input and $4.50 output for the Standard variant via Alibaba Cloud Int., pay-as-you-go with no markup. OpenRouter says four more providers exist; their names and prices were not confirmed here, so check the model page.

Prices vary by region and plan type. Press reports say Singapore pricing differs from Beijing and global deployments, and pay-as-you-go and Token Plan are billed differently. Token Plan subscription details were mostly found in third-party preview-period articles, so we do not cover them as official figures.

As a rough estimate, a request with 100,000 input tokens and 2,000 output tokens costs about $0.15 for input and $0.009 for output, roughly $0.16 in total. When you repeatedly send the same long prefix, cached input at $0.075 halves the input side. Alibaba reportedly claims 35% fewer tokens and about 60% lower average task cost than Qwen3.8-Max on its internal Work360 benchmark, but that is a vendor claim with no independent verification that we could find.

How to Use It — Calling the OpenAI-Compatible API from Python

The official model page says QwenCloud supports OpenAI-compatible, Anthropic-compatible and DashScope APIs. You can therefore use the OpenAI Python SDK and just change base_url and the model name. Issue an API key in the QwenCloud console and put it in an environment variable. The example below uses the Singapore endpoint.

# pip install openai
import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["QWEN_API_KEY"],  # API key issued in QwenCloud
    base_url="https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1",  # Singapore
)

resp = client.chat.completions.create(
    model="qwen4-max",
    messages=[
        {"role": "system", "content": "You summarize internal documents."},
        {"role": "user", "content": "Summarize these meeting notes in 3 lines.\n..."},
    ],
    # Only when you want deep reasoning (thinking)
    extra_body={"enable_thinking": True},
)
print(resp.choices[0].message.content)

To enable thinking, pass enable_thinking in extra_body, as the model page documents. Reasoning tokens are billed as output, and responses are usually slower, so leave it off for simple summarization or classification to save cost and time.

To use OpenRouter instead, use an OpenRouter API key and set the model to qwen/qwen4-max. OpenRouter's documentation lists supported parameters including temperature, top_p, top_k, max_tokens, tools, tool_choice, response_format, reasoning, reasoning_effort, seed and stop. It suits teams that switch between models with the same code or want consolidated billing.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["OPENROUTER_API_KEY"],
    base_url="https://openrouter.ai/api/v1",
)

resp = client.chat.completions.create(
    model="qwen/qwen4-max",
    messages=[{"role": "user", "content": "Summarize this contract in 5 bullet points.\n..."}],
    max_tokens=2000,
)
print(resp.choices[0].message.content)

At the time of checking, OpenRouter's model page stated that EU in-region routing is not available for this model. If you have data-location requirements, check this too.

Differences from Qwen3.8 and Other Vendors' Models

The previous top Qwen line is Qwen3.8-Max, reported at 2.4T total and about 95B active parameters with a 1M-token context. Qwen4-Max is reported at the same 2.4T scale but about 71B active. A lower active count suggests a design aimed at cheaper inference, but that is an inference from press-reported numbers, not something we confirmed officially.

ModelAvailabilityInputContextPrice (in/out per 1M)Local run
Qwen4-MaxAPI onlyText, image, video1M$1.50 / $4.50 (official)No
Qwen3.8-MaxMainly API; weights status reported inconsistentlyMostly text (press)1M (press)$2 / $6 (third-party listing)Unconfirmed
Qwen3.8-Omni-FlashAPI onlyText, image, audio, video1M$0.15 / $0.47No
Qwen3.8 27BOpen weightsMostly textVariesSelf-hosted (no API fee)Yes (needs GPU/VRAM)
Top GPT / Claude / Gemini modelsEach vendor's APIVariesVariesNot compared here (no verified official figures)No

For flagship models from other vendors (GPT, Claude, Gemini), we could not verify like-for-like official figures at the time of writing, so we make no claims about which is better or cheaper. The most reliable comparison is to run the same prompts from your own data through several models and measure quality, latency and unit cost.

If you need to process audio inexpensively, Qwen3.8-Omni-Flash has a far lower input price. Qwen4-Max is positioned as the top model for complex reasoning and very long contexts, so choose by use case.

Can You Run It Locally?

No. The weights are not released, so Qwen4-Max is available only through the API. Press reports say Alibaba will release smaller open-weight Qwen4 models later, but no date has been given. For now, if you need on-premises or local execution, pick a different open-weight model.

As covered in Qwen3.8 27B: requirements and VRAM, the open-weight Qwen3.8 family is an option for local runs. Check VRAM needs again when open Qwen4 models appear. A model at the 2.4T-parameter scale would far exceed what a personal PC can run even if its weights were released.

Caveats — Data Handling, Region and Preview

- Where data goes: With an API, your documents, images and video are sent to Alibaba's cloud, or to the underlying provider when using OpenRouter. If you handle personal or confidential data, review the terms and data-processing conditions first
- Region: The only endpoint we confirmed on the official page is Singapore (ap-southeast-1). Press reports say Beijing and global pricing is separate, so prices and terms differ by region
- EU in-region routing: Not available on OpenRouter at the time of checking
- Preview confusion: The Nov 11 limited preview uses the model ID qwen4-max-preview, separate from the GA qwen4-max. Third-party articles mention a preview shutdown on Nov 25, which we could not confirm officially, so do not use the preview ID in production
- Vendor-reported numbers: Benchmarks and the "60% lower cost" claim come from Alibaba itself; independent verification was not found
- Document inconsistencies: Max output tokens, batch inference and fine-tuning are described differently across sources. Check the console before production use

When to Choose It, and When Not To

- Choose it: you want to process many documents or videos together in a 1M-token context
- Choose it: you have OpenAI-compatible code and want to swap models to compare
- Choose it: your input includes images or video and text output is enough
- Skip it: you must keep weights in-house and run fully locally (look at open-weight models)
- Skip it: your data cannot leave the country (check regions and terms first)
- Skip it: you need audio or image generation (output is text only)

References (Primary Sources and Press)

We directly checked the official model page and the OpenRouter page. We could not directly read the body of Qwen's official blog post at the time of writing. CASRAI and MarkTechPost are secondary sources (MarkTechPost was not accessible when we checked).

What is Qwen4-Max?

It is the top-tier API model that Alibaba's Qwen team made generally available on November 16, 2026. It takes text, image and video input, outputs text, and has a 1M-token context. Weights are not released; it is API only.

How much does the Qwen4-Max API cost?

The official QwenCloud model page lists $1.50 per million input tokens, $4.50 per million output tokens and $0.075 for implicit cache input. OpenRouter via Alibaba Cloud Int. shows the same level. Prices may differ by region and plan, so confirm before use.

Can Qwen4-Max run locally?

No, because the weights are not released. Press reports say smaller open-weight Qwen4 models will follow, with no date. For local runs today, use an open-weight model such as Qwen3.8 27B.

How do I call it?

Use QwenCloud's OpenAI-compatible endpoint, or OpenRouter with the model name qwen/qwen4-max. With the OpenAI Python SDK, just swap base_url and the model name. Enable thinking via enable_thinking in extra_body.

Are the parameter counts officially confirmed?

Not at the time of writing. The 2.4T total / 71B active figures appear only in press coverage that attributes them to Alibaba, not on an official page we could read. Treat them as press-reported until the official blog confirms them.

Should I use qwen4-max-preview?

Use the GA model ID qwen4-max. The preview was a limited Token Plan release on November 11, and third-party articles mention a shutdown, so avoid it for production.

Qwen4-Max is an API model offering a 1M-token context and multimodal input at an official $1.50 input / $4.50 output per million tokens. Its weights are closed and some limits are described inconsistently across sources. Before adopting it, confirm region, data handling and current specs in the console, then measure quality and cost on your own data.

株式会社オブライト

Oflight Inc. is a Tokyo-based IT company providing software development, digital signage, web development, network systems, and IoT/AR/VR technology services.

Related free tools (no sign-up, instant results)

Feel free to contact us

Contact Us