Qwen4-Max Is Out of Preview: 2.4T Parameters, Still No Weights

A glass server tower glowing violet among data center racks, next to a wooden desk with a small monitor and keyboard, with a Qwen logo badge in the corner

Alibaba made Qwen4-Max generally available on 16 November 2026, five days after the Qwen4-Max preview. The model ID is now qwen4-max, it has a per-token price, and the Qwen team published its size: a 2.4 trillion parameter mixture-of-experts with 71B parameters active per token (Qwen blog, 2026). Short answer for local users: you still cannot run it. Alibaba has published no weights, and the Model Studio page lists fine-tuning and model deployment as not supported (Alibaba Cloud Model Studio, 2026).

What changed from the preview?

The preview ran on qwen4-max-preview through Token Plan only, with no price and no parameter count. The general release fills most of those gaps:

Preview (11 November)General availability (16 November)
Model IDqwen4-max-previewqwen4-max
AccessToken Plan onlyModel Studio API, with per-token pricing
PriceNot published$1.50 input, $4.50 output per million tokens (Singapore, international)
ParametersNot published2.4T backbone, 71B active, plus 328B N-gram embedding parameters
WeightsNoneNone

Sources: Model Studio, Qwen4-Max, Model Studio pricing and the Qwen blog, all dated 16 November 2026.

What does the official spec say?

The Model Studio page, updated on 16 November 2026, is the reference:

SpecValue (Alibaba Cloud)
Model IDqwen4-max
InputText, images and video
OutputText
Context length1,000,000 tokens
Maximum input991,808 tokens without thinking, 983,616 with thinking
Maximum output131,072 tokens without thinking; 262,144 for thinking and answer combined
ThinkingOn by default, reasoning_effort defaults to high
API compatibilityOpenAI Chat Completions, OpenAI Responses, Anthropic Messages, DashScope
Fine-tuningNot supported
Model deploymentNot supported
Batch inferenceNot supported

The price is lower in Beijing and in the Global regions (Hong Kong, Frankfurt, Virginia and Japan): $1.413 input and $4.24 output per million tokens, per the pricing page. Cache reads cost $0.075 per million tokens in Singapore.

"Model deployment: not supported" matters for anyone hoping to host the model privately. You can only call it on Alibaba's servers.

How good is it?

Bloomberg covered the launch as Alibaba's "riposte to US rivals" (Bloomberg, 2026). The Qwen blog's chart compares Qwen4-Max with Anthropic, OpenAI and Moonshot models across "coding, work, and long-horizon tasks". It calls the model "a significant step up from Qwen3.8-Max". Every score in that chart comes from Alibaba's own evaluation.

The first independent number comes from Artificial Analysis, which gives Qwen4-Max a score of 43 on its Intelligence Index, and lists the same $1.50 and $4.50 prices (Artificial Analysis, 2026). We do not rank cloud models on ModelFit, so Qwen4-Max gets no model page and no entry in the open dataset.

Can a 512GB Mac Studio run Qwen4-Max?

No, even if the weights shipped tomorrow. With a parameter count now published, the gap can at last be sized. ModelFit's rule of thumb for 4-bit quantization is about 0.6GB per billion parameters (how much VRAM for LLMs). For the 2.4T backbone alone, that is about 1,440GB est., before the 328B N-gram embedding parameters and the KV cache.

MachineMemoryShare of the Q4 backbone (est.)
Mac Studio M5 Ultra512GBAbout a third
Two DGX Spark256GBUnder a fifth

So Qwen4-Max would need roughly three 512GB Mac Studios at 4-bit, the same class as the previous flagship. Qwen3.8-Max is also a 2.4T model, and it is still cloud-only.

Will smaller Qwen4 models be open-weight?

Alibaba says yes, with no date. The Qwen blog closes on this: "Max is only the first look at Qwen4. Smaller models in the family are coming to open source." It names no size, no license and no release date. As of 16 November 2026, nothing named Qwen4 exists in the Qwen organisation on Hugging Face, per the Hugging Face API.

Until they arrive, the closest local preview of the architecture is Qwen3.8-Flash-Next, which Qwen describes as a preview of the architecture behind Qwen4. It runs from a 128GB Mac, or on smaller Macs with SSD streaming (Strata on a Mac). For a dense model on a 24GB GPU, Qwen3.8-27B is the current pick. The ModelFit wizard or npx @modelfit/cli@latest matches a model to your hardware.

FAQ

Can you run Qwen4-Max locally?

No. Alibaba has published no weights, and Model Studio lists model deployment and fine-tuning as not supported. Qwen4-Max runs only through Alibaba's API and Qwen apps.

How many parameters does Qwen4-Max have?

2.4 trillion backbone parameters with 71B active per token, plus 328B N-gram embedding parameters, according to the Qwen team's launch post.

How much does the Qwen4-Max API cost?

$1.50 per million input tokens and $4.50 per million output tokens in the Singapore region, and $1.413 and $4.24 in Beijing and the Global regions, per Alibaba Cloud's pricing page.

What is the Qwen4-Max context window?

1,000,000 tokens. The maximum input is 991,808 tokens without thinking and 983,616 with thinking.

Will Qwen4 have open weights?

Alibaba says smaller Qwen4 models are coming to open source, without a date or sizes. We will add them to the open dataset once weights ship with support in llama.cpp or mlx-lm.

Sources

  • Alibaba Cloud Model Studio, Qwen4-Max (updated 16 November 2026): https://www.alibabacloud.com/help/en/model-studio/qwen4-max
  • Alibaba Cloud Model Studio, model pricing: https://www.alibabacloud.com/help/en/model-studio/model-pricing
  • Qwen team, Qwen4-Max: Your New Teammate Has Arrived (16 November 2026): https://qwen.ai/blog?id=qwen4-max
  • Bloomberg, Alibaba Unveils New Flagship AI Model in Riposte to US Rivals (16 November 2026): https://www.bloomberg.com/news/articles/2026-11-16/alibaba-unveils-new-flagship-ai-model-in-riposte-to-us-rivals
  • Artificial Analysis, Qwen4 Max: https://artificialanalysis.ai/models/qwen4-max
  • Hugging Face API, Qwen organisation models, read 16 November 2026: https://huggingface.co/api/models?author=Qwen&search=Qwen4
What hardware runs this?

Match this model to a machine that can run it: by RAM tier for Apple Silicon, or by VRAM for an NVIDIA GPU.

See how this changes your recommendation

Run the wizard

Have questions? Reach out on X/Twitter