Qwen 4: Release Date, Qwen4-Max Preview, Model Sizes and Running It Locally

Qwen 4 has no release date. Alibaba said on 22 September 2026 that Qwen 4 is in training and gave no sizes or licence. The only Qwen 4 model you can use is qwen4-max-preview, which opened on 11 November on Alibaba Cloud's Token Plan subscription and retires after 25 November. No Qwen 4 weights are downloadable, and no tier has been confirmed as open weight.

Search interest in Qwen 4 has outrun the facts. Release dates, a 27B model and open weights are all being discussed as if they were settled, and none of them is. This page sorts what Alibaba has published from what has only been reported, and turns the rest into a hardware plan you can use whichever way the announcement goes.

Everything here was checked against Alibaba Cloud’s documentation and press room and Qwen’s Hugging Face organisation on 13 November 2026. We will update it when Alibaba announces Qwen 4 sizes or weights.

Qwen 4 release date: what Alibaba has said

Alibaba has made two public statements about Qwen 4, and neither contains a release date.

1. 22 September 2026, Apsara Conference, Hangzhou. Alibaba’s roadmap press release says its next-generation model, Qwen 4, “is currently in training”. The same release projects that the later Qwen 4.5 and Qwen 5 series will scale to 5 to 10 trillion parameters. That figure is a target for those future generations, not a size for Qwen 4.
2. 11 November 2026, Model Studio documentation. A new qwen4-max-preview page describes a “next-generation flagship multimodal model” aimed at coding, office work, front-end design and agent tasks, available for a fixed preview window.

Between those two dates, release dates spread through social posts and aggregator sites, including claims in late October that Qwen 4 would ship the following week. They did not hold. The preview that did arrive is a hosted, time-limited model rather than a general release, so the question of when Qwen 4 itself ships is still open.

Qwen’s own pattern offers a rough guide, not a forecast. For the previous generation, Qwen published its open Qwen3.8 checkpoints during August 2026, with the 2.4-trillion-parameter base of the hosted Qwen3.8-Max among them. Whether Qwen 4 follows a similar path is something Alibaba has not addressed.

Qwen 4 confirmed, reported and unknown

Here is the state of each claim on 13 November 2026.

The last row deserves weight. The Qwen3.8-Flash-Next model card describes the model as an “experimental preview of the architecture that will underpin Qwen4”. It pairs a mixture of 512 experts and block-sparse attention with a large n-gram embedding table that can sit outside GPU memory. That is a statement about design direction, not a promise that any Qwen 4 checkpoint will match Flash-Next in size.

What is Qwen4-Max preview?

Qwen4-Max preview is the first Qwen 4 model anyone outside Alibaba can call. Its documented specification:

Two behaviours catch existing integrations out. In the Chat Completions format, max_tokens means the same as max_completion_tokens and covers the thinking as well as the answer, so a cap sized for answers alone can cut a response off mid-reasoning at the default xhigh effort. And Chat Completions keeps prior reasoning by default (preserve_thinking), which expects earlier thinking to come back in the reasoning_content field rather than pasted into content. Do not set reasoning_effort and thinking_budget in the same request.

The “Max” name matters for planning. In the Qwen3.8 generation, Max was the largest tier, and the open checkpoint behind it has 2.4 trillion parameters. Even if a Qwen4-Max checkpoint were ever published, it would be cluster hardware, not a workstation model.

How to try the Qwen4-Max preview

1. Get a Token Plan. In Alibaba Cloud Model Studio, subscribe to Token Plan Personal or Token Plan Team. Standard pay-as-you-go keys and Coding Plan keys cannot call the preview.
2. Create a Token Plan key. These begin with sk-sp- and are not interchangeable with ordinary Model Studio keys. Keys, base URLs and plans must be used in matching pairs.
3. Point your client at the Token Plan endpoint. For international accounts the OpenAI-compatible base URL is https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1 and the Anthropic-compatible one ends in /apps/anthropic. The China-site documentation lists a cn-beijing host instead.
4. Set the model to qwen4-max-preview.

curl https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/chat/completions \
  -H "Authorization: Bearer $TOKEN_PLAN_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen4-max-preview",
    "reasoning_effort": "medium",
    "max_tokens": 32768,
    "messages": [{"role": "user", "content": "Summarise the trade-offs of block-sparse attention."}]
  }'

Built-in model tools work only through the Responses format; Chat Completions cannot invoke them.

Token Plan bills in monthly Credits rather than per token. During the promotion, Alibaba’s Token Plan page says qwen4-max-preview consumes Credits at 20% of the standard rate, and at 8% between 22:00 and 08:00 UTC+8. Subscription prices and new-customer discounts change often, so read them on the Token Plan Personal overview on the day you buy.

What leaves your network

Calling the preview means sending prompts, attachments and tool results to Alibaba Cloud. The Token Plan Personal page lists Singapore as the only region and Global as the deployment mode, and it notes that prompts and outputs involve cross-border data transfer. It does not say how long requests are kept.

For a personal test of public material, that may be fine. For source code, customer records or regulated documents, the preview is a different decision: the model runs on hardware you do not control, under terms you should read with your data protection lead, and it disappears after 25 November. Use it to judge capability with synthetic or public prompts, not as somewhere to send production data. If the reason you follow Qwen is that its open checkpoints can run inside your own perimeter, the preview tells you about quality, not about deployment.

Qwen 4 model sizes: what we know

Very little. Alibaba’s Qwen 4 statements contain no parameter counts, and the preview page omits them too.

The tier names in circulation, Max, Plus, Flash and 27B, come from reporting on the Apsara keynote slides. They are plausible, because they mirror names Qwen already uses: Max and Flash exist in the Qwen3.8 line, Plus in the Qwen3.7 line, and 27B is the dense Qwen3.8 size. Plausible is not confirmed. A “Qwen 4 27B” could arrive as a dense 27-billion-parameter model, as a mixture-of-experts model with a different total, or not at all.

Searchers also ask about “Qwen 4.0”. The only official use of that string is the billing alias qwen4.0-max-preview, which refers to the same preview model.

On Hugging Face, Qwen’s organisation held six Qwen3.8 repositories and no Qwen 4 repositories on 13 November 2026. Anything labelled Qwen 4 on a community account is somebody else’s upload, a fine-tune or a mislabelled file, and should be treated with the caution you would give any unknown binary.

Will Qwen 4 be open weight?

Nobody outside Alibaba knows. The press release does not mention open weights or licences for Qwen 4, and the preview is a hosted service.

The Qwen3.8 generation shows why “Qwen is open source” is too simple an assumption:

• Qwen3.8-27B shipped under Apache 2.0.
• Qwen3.8-Flash-Next shipped under the Qwen Community License 1.0, which adds conditions for very large products and for companies selling model or AI-assistant services.
• Qwen3.8-2.4T-A95B, the base of the hosted Qwen3.8-Max, shipped under its own Qwen3.8-Max License.
• Qwen3.8-Flash and Qwen3.8-Omni-Flash stayed API-only.

So even if Qwen 4 weights appear, expect different terms per tier. Our guide to open-weight licence terms for enterprise use explains how to read these licences before you commit hardware or product plans to a model.

What running Qwen 4 locally will take

This section is a planning aid, not a specification. Every figure below is an estimate from parameter counts and from the Qwen3.8 models a Qwen 4 tier might resemble, worked out with the method of our LLM VRAM calculator: weights are parameters times bits per weight divided by eight, add the KV cache for the context you keep, then divide by 0.9 so the runtime has headroom (0.8 for unified memory shared with the operating system).

The middle columns borrow real file sizes from the Qwen3.8 equivalents: Qwen3.8-27B is 55.6 GB in BF16 and 30.9 GB in FP8, with Unsloth’s 4-bit GGUF files at 14.4 to 16.6 GB; Flash-Next is about 186 GB in FP8, with 4-bit community files of 93.7 to 120 GB; and the 2.4T model’s FP8 release is about 2,496 GB.

Working through the 27B-class case

Take a 16.6 GB 4-bit file. If a Qwen 4 27B kept Qwen3.8-27B’s attention layout, a token of 16-bit KV cache would cost 64 KiB. Then:

• 32,768 tokens of context: 16.6 + 2.1 = 18.7 GB, divided by 0.9 gives about 21 GB, so a 24 GB card such as an RTX 4090 works.
• 131,072 tokens: 16.6 + 8.6 = 25.2 GB, or about 28 GB to provision, which points to a 32 GB card such as an RTX 5090.
• 8-bit weights at 131,072 tokens: 30.9 + 8.6 = 39.5 GB, or about 44 GB, which needs a 48 GB card.

If Qwen 4 changes the attention design, the cache line moves, so re-run the numbers in the calculator once a config file exists.

Working through the Flash-class case

A Flash-Next-style model keeps every expert and its n-gram table loaded even though each token touches only about 6 billion parameters. A 93.7 GB 4-bit file fits the roughly 102 GB usable on a 128 GB DGX Spark, Ryzen AI Max+ 395 system or Mac, with room for long context because Flash-Next’s sparse attention caches little per token. Speed then depends on memory bandwidth more than capacity; our tokens-per-second calculator shows the ceiling for each machine, and our comparison of 128 GB desk-side systems weighs them against each other.

For a shared server, the 8-bit build would span several data-centre GPUs; our GPU buying guide for local models covers those cards.

Buying hardware before the announcement

If you must buy now, buy for a range. A 32 GB GPU covers any 27B-class model at 4-bit with generous context. A 128 GB unified-memory machine covers both the 27B class at full precision and a Flash-Next-sized model at 4-bit. Anything larger is a server decision best made after sizes are public.

Qwen 4 vs Qwen 3.8: should you wait?

For most teams, no. Waiting buys nothing while the release date, sizes and licence are all unknown, and the work that makes a model useful does not depend on its version number.

Reasons to start on Qwen3.8 now:

• The models exist. Qwen3.8-27B under Apache 2.0 fits a single GPU, and Flash-Next, the architecture preview for Qwen 4, fits a 128 GB machine. Our Qwen3.8 hardware and file-size guide lists every build with its memory needs.
• The evaluation carries over. A test set of your own prompts, documents and tool calls, with pass criteria, is what will tell you whether Qwen 4 is worth switching to. Building it on Qwen3.8 means you can score Qwen 4 the week it appears.
• The swap can be cheap. Serve the model behind an OpenAI-compatible endpoint, as vLLM, SGLang, llama.cpp and Ollama all provide, and changing the model later becomes a configuration change rather than an application rewrite. Our vLLM, Ollama and llama.cpp comparison helps choose the server.
• The preview is temporary. qwen4-max-preview retires after 25 November, and Alibaba promises no successor date, so nothing should be built on it.

Reasons to hold back:

• A large hardware purchase that only makes sense for one specific Qwen 4 size.
• A licence-sensitive product, where you need to see Qwen 4’s terms before choosing between Qwen and another family.
• A benchmark-driven choice between frontier hosted models, where the preview itself is the thing to test.

Whichever way you go, plan the upgrade path in advance: how a new model is approved, tested against the old one and rolled out. Our model governance and upgrade guide sets out that process.

What to run today while Qwen 4 trains

For software teams, the coding picture is broader than Qwen. Our ranking of local models for coding compares Qwen3.8-27B and the Qwen3-Coder line with Devstral, gpt-oss, GLM and others by memory tier. For organisation-wide rollouts across several Qwen sizes, see our enterprise guide to running Qwen on premises. If you are weighing other large open models at the same time, our Kimi K3.1 Pro local hardware guide sizes up a rival open-weight family released this month.

How VDF AI fits

VDF AI does not run, host or bundle Qwen 4, the Qwen4-Max preview or any Qwen model. It is the layer around whichever model you choose: an on-premises or private-cloud platform for governed AI agents, multi-agent workflows in VDF AI Networks, retrieval over company documents that respects permissions, role-based access control on every plan and audit logs of who asked what.

The model question in this article, try a hosted preview or wait for weights you can run yourself, is the one VDF AI Router is built around. It registers Ollama and other OpenAI-compatible on-premises endpoints, such as a vLLM server running Qwen3.8-27B today, next to the cloud models your policy allows, and it prefers the local ones when they are available. Air-gap mode restricts traffic to local models entirely. If an open Qwen 4 checkpoint appears, it becomes one more endpoint to register and test, with the applications that call the router unchanged. Production self-hosting runs on Linux, as the infrastructure requirements explain.