Qwen 4 Release Date: Not Out Yet, What We Know and Qwen 3.8 Today

No. As of 9 November 2026, Qwen 4 is not released. The official Qwen account on Hugging Face lists no Qwen4 repository, Alibaba has published no Qwen 4 model ID, price or specification, and no release date has been given. Alibaba teased something on 6 November. The newest models you can actually call or download are the Qwen3.8 family from August.

Is Qwen 4 out yet?

No. We checked the Qwen organisation page on Hugging Face on 9 November 2026. The first ten repositories it listed were all Qwen3.8 or Qwen3.6 models: the 2.4T-A95B, Flash-Next and 27B checkpoints, each with an FP8 build, plus Qwen3.6 35B-A3B and 27B. None used a Qwen4 name.

One caveat applies. The page showed 10 of 469 repositories, so that check alone cannot rule out a name further down the list. It does match what the news says: Alibaba has teased a successor but has not shipped it.

You should also not trust any page that gives Qwen 4 parameter counts, context length, benchmark scores or a price. Nothing Alibaba has published supports them.

What has Alibaba actually said about Qwen 4?

Three things are on the record.

The teaser. On Friday 6 November, the official @Alibaba_Qwen account on X posted a poster with the English line "The becoming begins." The poster also carries a classical Chinese line that breaks off before a name. Asia Times describes it as the opening of a Zhuangzi passage, where a fish named Kun becomes the giant bird Peng. The post does not say "Qwen 4". Asia Times notes that it "announces neither a release date nor a technical specification". Reading it as a Qwen 4 teaser is a reasonable inference, not an announcement.

The architecture preview. The Hugging Face card for Qwen3.8-Flash-Next calls itself an "experimental preview of the architecture that will underpin Qwen4". That is a primary source. It tells you Alibaba plans a Qwen4 and that Flash-Next previews the design. It tells you nothing about timing.

Training status. Alibaba's own press release dated 22 September 2026, from its Apsara Conference, says Qwen 4 is "currently in training". It gives no release date and no specifications for Qwen 4. Asia Times adds that CEO Eddie Wu praised the existing Qwen 27B model and promised better models for running locally.

Not confirmed. Reading the 6 November teaser as a Qwen 4 announcement is our inference. Alibaba has not named the model or given a date.

When will Qwen 4 be released?

Nobody outside Alibaba knows, and Alibaba has not said. There is no date, no window and no pricing.

Here is what we can reason from, labelled as our inference. A public teaser suggests a launch is nearer, so a release in the coming weeks is plausible. That is a pattern, not a commitment. Plan for "soon but unknown", not for a date.

The practical point is that the date matters less than your readiness. If a release lands on a Tuesday, you want to test it by Friday. That depends on work you can do now, covered in the last sections below.

Also expect a family, not one model. Qwen3.8 shipped as several checkpoints with different sizes and different licences. If Qwen 4 follows that pattern, the model you want and the licence you can accept may not be the one in the headline.

What is the Qwen3.8 line-up today?

The newest general-purpose generation on Hugging Face is Qwen3.8, released in August 2026. Hub pages show update dates, not release dates, so we give the month only. These are the three open-weight checkpoints.

Two details matter for buyers. First, the licences differ by size. The 27B card reads "License: apache-2.0", and the licence file on the repository is the full Apache 2.0 text. The 2.4T card names a custom licence, "qwen3.8-max". Read that licence before you plan around it. We cover the larger model in our Qwen3.8 2.4T deep dive.

Second, the 2.4T card says "thinking cannot be disabled" and that multimodal input is not supported. The 27B card says thinking is on by default and can be turned off. The hosted Qwen3.8-Max, which is a separate API model, is not the same as the open 2.4T checkpoint.

For running the 27B yourself, our Qwen3.8-27B self-hosting guide covers memory and serving.

How much does the Qwen API cost?

Alibaba Cloud Model Studio lists Qwen3.8 prices on its pricing page. The International (Singapore) table, marked "Last Updated: Nov 07, 2026", gives these standard prices per million tokens in US dollars.

Notes from the page. Output for the max and 2.4t rows covers "Chain of thought + answer", so thinking tokens are billed as output. Each of these four models carries a free quota of 1 million tokens, valid for 90 days, and the page says the free quota applies "only in Singapore". Current promotions are in the console, not on the page.

On caching, the page gives an example rule: explicit cache creation is billed at 125% of the standard input price, and cache hits at 10%. Alibaba's Context Cache page excludes qwen3.8-flash from those general ratios and points to the console for its prices. The input prices in the table do not include cache prices. On batch, the page says batch calls cost 50% of the real-time price, but the Singapore rows for these four models carry no batch label. A batch label appears on the Beijing rows only. Do not assume batch pricing in Singapore.

Alibaba also lists Global deployment prices for qwen3.8-max of about $1.65 input and $4.95 output per million tokens. Our table and examples use Singapore's International deployment scope.

A worked example, with our assumptions. Take 10,000 requests a month, each with 3,000 input tokens and 500 output tokens. That is 30 million input and 5 million output tokens. This ignores caching, and for the Max and 2.4T rows it understates output, because thinking tokens count.

• qwen3.8-flash: 30 x $0.15 = $4.50, plus 5 x $0.47 = $2.35, so $6.85.
• qwen3.8-27b: 30 x $0.50 = $15, plus 5 x $3 = $15, so $30.
• qwen3.8-max: 30 x $2 = $60, plus 5 x $6 = $30, so $90.

Those are our arithmetic on Alibaba's rates. The spread is more than tenfold from Flash to Max. Which model is good enough for your task decides the bill far more than the rate card does. For a method that prices tasks, not tokens, see our post on per-task cost.

The Flash-Next card describes hosted Qwen3.8-Flash as "the official version based on Qwen3.8-Flash-Next with more production features". That does not establish identical behaviour between the API and the downloaded checkpoint.

How do you get a Qwen API key?

Create it in the Alibaba Cloud Model Studio console. The key page says to use an Alibaba Cloud account, or a RAM user with the right permissions, choose your region in the upper right, and click Create API Key. The key and its API Host appear once. Copy them immediately, because you cannot view the full key again.

Keys are tied to a region. Alibaba's page says each region "has its own endpoint, API keys, and model list" and that they cannot be used across regions. A Singapore key will not work against Beijing endpoints. This matters for the free quota, which is Singapore only, and for data location. The page suggests storing the key in an environment variable named DASHSCOPE_API_KEY.

We did not find the Singapore OpenAI-compatible base URL on that page. Take it from the API Host shown when you create the key, not from a blog post, ours included.

Which Qwen model should you use today?

Our view, by situation. These are our inference from the cards and prices above, not Alibaba guidance.

You need an API with the lowest cost. Start with qwen3.8-flash. At $0.15 and $0.47 it is the cheapest row, and your evals decide if it is good enough.

You need to run it yourself on one machine. Use Qwen3.8-27B. It is Apache-2.0, and the Ollama library page lists a 27b build at 18GB with a 256K context, labelled with tools, vision and thinking. That figure is a download size from Ollama's page, not a memory requirement. Our self-hosting guide explains the gap. No Qwen card we read states a GPU memory requirement.

You need the strongest Qwen. Use the hosted qwen3.8-max, at $2 and $6. Its published benchmark comparisons are for the hosted model, so do not attribute them to the open 2.4T weights.

You want to wait for Qwen 4. Do not hold production work for an unannounced model. Build on Qwen3.8 now and make the swap cheap, as the next section describes.

Benchmark figures on the Qwen cards are the vendor's own, not independently run. Test on your own tasks.

How do you prepare for Qwen 4 release day?

Treat the release like any model swap. Four steps, in order.

1. Keep the model ID in one place. A single config value, not strings scattered across services. Release day then means changing one line, and rolling it back means changing it again.
2. Keep an eval set. Collect 50 to 100 real tasks with the outputs you accept. Run every candidate against it. A model that wins public benchmarks can still lose on your data.
3. Check the licence variant before you test. Qwen3.8 splits licences by size: Apache-2.0 for the 27B, a community licence for Flash-Next, a custom one for the 2.4T. A Qwen 4 family may split again, and the model you want may not be the permissive one.
4. Price it on your own token mix. A new tokenizer or a new thinking default can move cost per task without changing the rate card. Our piece on hidden reasoning-token costs explains why.

Then wait for primary sources: a Hugging Face model card, an Alibaba Cloud pricing row. Do not act on a screenshot or a leaderboard entry for a model with no card.

If you build agents on Qwen, also read how tool calling behaves today in our guide to how AI agents call Qwen 3.8. Tool-call formats are the part most likely to change between generations.

Quick answers

Is Qwen 4 released? No. Teased on 6 November, with no model ID, date or price.

Who says so? The Hugging Face organisation page on 9 November, the Qwen3.8-Flash-Next card, and Asia Times for the teaser.

What is the latest Qwen I can use? Qwen3.8, from August 2026, open weights and hosted API.

How much is the Qwen API? From $0.15 and $0.47 per million tokens for qwen3.8-flash up to $2 and $6 for qwen3.8-max, on Alibaba's Singapore page updated 7 November.

Is Qwen free? The open weights are free to download under their licences. The API gives a 1 million token free quota for 90 days in Singapore.

Where Swfte sits

If you are waiting on a model release, the useful thing is a cheap way to switch. Swfte Connect offers one API across many providers, with smart routing that optimises cost and availability. Connect has per-workspace routing rules with a mandatory fallback chain, so a request can fall to another model when the first one fails.

When a workspace reaches its usage cap, Connect moves its requests to a cheaper model. Automatic downgrade can reduce further spending. A cheaper model still incurs charges, so set a request limit for each evaluation run.

You can also bring your own provider key. We have not verified that a Qwen model of any size is available in your workspace, so check the catalogue. A model outside the catalogue can be added as bring-your-own-key with a base URL, so your own OpenAI-compatible endpoint (for example a self-hosted Qwen3.8 on vLLM) can sit behind Connect with routing, fallback and usage caps. Test the specific checkpoint before you rely on it.

The decision rule is short. Build on Qwen3.8 today, keep the model ID in one place, and treat Qwen 4 as a test to run when a model card exists. We cover routing in more depth in intelligent LLM routing.

---

Related: Qwen3.8-27B self-hosting guide, Qwen3.8 2.4T open weights deep dive, How AI agents call Qwen 3.8, Claude Opus 5.5 per-task pricing.

Sources: Hugging Face, Qwen organisation page · Hugging Face, Qwen3.8-27B model card · Hugging Face, Qwen3.8-Flash-Next model card · Hugging Face, Qwen3.8-2.4T-A95B model card · Alibaba Cloud, Model Studio pricing · Alibaba Cloud, Press release, 22 September 2026 · Alibaba Cloud, Context cache · Alibaba Cloud, Get an API key · Asia Times, What Alibaba left blank about Qwen4 · Ollama, qwen3.8 library page · accessed 9 November 2026.

Every figure was re-read on its primary page on 9 November 2026. Qwen benchmark claims are vendor-reported and not independently reproduced. Worked cost examples are our arithmetic on assumed workloads, not measurements. Qwen 4 has no published specification, so none is given.