Qwen 4: Release Date, Qwen4-Max Preview, and What's Confirmed (2026)

Qwen4-Max is in preview as of November 11. The rest of Qwen 4 still has no date. Here's what's actually known, separated from the leaks.

Search for “Qwen 4 release date” and you’ll find confident headlines naming a month. The September ones trace back to one July leak with no documentation, and September passed without a launch. Meanwhile the one thing Alibaba has actually confirmed, that the Qwen 4 architecture is already public inside a downloadable model, gets far less attention than it deserves.

This post keeps those two categories separate. It will be updated the day Qwen 4 ships.

Update, September 22, 2026: At the Apsara Conference in Hangzhou, Qwen project lead Liu Dayiheng said Qwen 4 is in training on a new-generation architecture and will be released "very soon," and that Qwen 4.5 and Qwen 5 will scale to 5 to 10 trillion parameters. No date, tiers, or specs were given. The sections below are updated where that changes them.

Update, October 26, 2026: Qwen 4 still hasn't shipped and Alibaba hasn't given a date. An unreleased model codenamed Kalmia is being tested on Arena and testers are attributing it to Qwen 4, but nothing is confirmed; details in the evidence section below. The prediction market sits at 83 percent for a launch before November.

Update, November 7, 2026: On November 6 the official Qwen account on X posted a teaser: a poster of birds over a dark sea, the flock shaped like a fish, with the line “The becoming begins.” and the Chinese 北冥有鱼，其名为 _, the opening of Zhuangzi’s Kun passage with the last character left blank. It names no model and no date, but it is the first official signal since September that a launch is close, and attendees at Alibaba Cloud’s October 29 event in Amsterdam have reported an executive saying “Qwen4 next week,” which Alibaba has not confirmed. November arrived without a launch, so the 83 percent market call missed. Still no weights, no API model IDs, and no announcement as of this update.

Update, November 8, 2026: On November 7 Alibaba Cloud announced a stealth model called Qoduck inside its Qoder coding IDE, origin undisclosed, free daily for individual users for a limited time, with no public API. Alibaba has not said what it is, and testers are treating it as a Qwen 4 preview based on tokenizer and output behavior. Together with the November 6 teaser, it points to a launch within days, but Qwen 4 itself is still unreleased: no weights, no model IDs, no pricing.

Update, November 11, 2026: Qwen4-Max is in preview. At 10:00 Beijing time on November 11, the official Qwen account posted "Qwen4 is coming." and Alibaba listed Qwen4-Max-Preview (model ID qwen4-max-preview) on its Token Plan subscription, with early access in QwenWork, Qoder, the Qwen app, and QwenStudio. Alibaba's docs give it a 1M-token context, up to 393,216 output tokens, text, image, and video input, and thinking on by default. It is available through Token Plan only, with no pay-as-you-go API price, and the preview runs through November 25, after which Alibaba says the preview model will be retired. There are no published benchmarks, no open weights, and no date for the full release or for Plus, Flash, and the 27B.

TL;DR

• Qwen4-Max is in preview as of November 11, 2026, on Alibaba's Token Plan subscription only. The full Qwen 4 release has no announced date, and Alibaba's launch post says only "Qwen4 is coming."
• What Alibaba has confirmed: Qwen 3.8-Flash-Next, released August 28, is "an experimental preview of the architecture that will underpin Qwen4," shipped "ahead of the complete Qwen 4 rollout"
• The preview model is qwen4-max-preview: 1M-token context, text, image, and video input, thinking on by default. The preview window runs November 11 to November 25, and Alibaba says the preview model is retired after that
• The architecture signal is real: lean-activation MoE, a large n-gram memory layer, multimodal by default, 1M context
• The Qwen 3.8 rollout ran preview to flagship to open weights in about a month; Qwen 4 will likely follow the same staged pattern
• Alibaba has named four tiers, Qwen 4 Max, Plus, Flash, and a 27B open-weights model. Only Max has shipped, and only as a preview. Plus, Flash, and the 27B have no specs, pricing, or date, and there is no Qwen 4 Coder

Qwen 4 status at a glance

What Alibaba has actually confirmed

One thing, and it's substantial. On August 28, Alibaba released Qwen 3.8-Flash-Next and described it in the model card as "an experimental preview of the architecture that will underpin Qwen4," released so developers can prepare ahead of the full rollout. That's an unusual level of roadmap disclosure from a lab that usually announces by shipping.

So the Qwen 4 architecture is partly public today: a sparse MoE with 125B total parameters and only 6B active per token, a 51B-parameter n-gram embedding component that trades computation for memorized lookup, native image and video input, and a 262K context extensible to 1M. Whether the flagship Qwen 4 scales those choices up or refines them is unknown, but the direction is on the record: leaner activation, a memory layer, multimodal by default.

Everything else in this post is inference from that, from Alibaba's release cadence, or from rumors, and it's labeled accordingly.

The release date: what the evidence actually says

There is no official date for the full release, but the first Qwen 4 model is now in preview. Here's the weight of each signal, newest first.

The Max preview: On November 11, Alibaba put Qwen4-Max-Preview on its Token Plan with a fixed two-week window that ends November 25, and says the preview model will be retired when the window closes. A dated preview with a retirement date is the strongest signal so far that the general release is planned for around the end of that window, though Alibaba has not said so.

The Arena sighting: Since around October 22, an unreleased model under the codename Kalmia has been appearing in Arena's anonymous testing, and testers on X and Reddit have been attributing it to Qwen 4 based on its output style. Arena's own FAQ says pre-release models run under codenames and that a codename confirms nothing, and neither Alibaba nor the Qwen team has commented. Treat it as a rumor that fits the timing. If it is Qwen 4, a release usually follows an Arena run by days to a few weeks, not months.

The engines: vLLM is building Qwen 4 support before the weights exist. An open pull request titled "Add qwen4 fuse op" implements the sparse-attention, HyperConnection, and n-gram kernels from Flash-Next under the Qwen 4 name, a merged change switched vLLM to the upstream Qwen4-Exp model configs from the Transformers library, and two PLE-offload changes landed in September. That is the same order of events that preceded Qwen 3.8: engine support first, weights after. It is not a date, but it means day-one serving support is being written now, which is the strongest hint yet that the weights are close.

The Apsara keynote: On September 22, Qwen project lead Liu Dayiheng told the Apsara Conference that Qwen 4 is in training on a new-generation architecture and will be released "very soon," with Qwen 4.5 and Qwen 5 to follow at 5 to 10 trillion parameters. It's the first time Alibaba has put Qwen 4 on a public stage, and "very soon" from a lab that ships every two to three weeks is the strongest timing signal so far. It is still not a date.

The September rumor: A July 20 leak, attributed to a YouTube channel, claimed a September 2026 launch and described experimental codenames with 3D-generation abilities. No documentation, no Alibaba confirmation, and the same leak's other claims haven't been independently verified. September has now passed without a launch, so treat it as a guess that got repeated.

The market: A public prediction market on the Qwen 4 release date priced a launch before October 1 at 44 percent in early September, and that window closed without one. As of October 23 it prices a launch before November 1 at 83 percent. That’s the crowd saying October is more likely than not, and not certain.

The cadence: Alibaba's recent rhythm is the strongest evidence. The Qwen 3.8-Max rollout went preview on July 19, launch on August 3, open weights around August 12-14, then Flash-Next on August 28, something shipping every two to three weeks. A preview explicitly framed as Qwen 4 preparation suggests the flagship isn’t far behind, and Alibaba used its Apsara Conference on September 22 to name the tiers without shipping them. That's pattern-matching, not a date.

The practical read: plan for a fall launch, don't plan around a specific week, and expect the same staged rollout as 3.8, preview or announcement first, API access, then open weights for some tiers.

What Qwen 4 will probably look like

If Flash-Next is the blueprint, and Alibaba says it is, expect three things. Much lower active-parameter counts relative to total size, which means cheaper serving per token and the same MoE memory rule as always: every parameter still needs VRAM. The n-gram memory layer, the most novel bet, which trades transformer compute for lookup and would matter for both cost and how the model behaves on factual recall. And multimodality as the default rather than a variant.

What the leak claims, 3D generation, gaming and architecture applications, is unverified and not something the confirmed architecture speaks to either way. Benchmarks don't exist yet, obviously; when they arrive they'll be vendor-reported first, and the Qwen 3.8 benchmarks tracker shows how long independent replication takes for this family, typically weeks.

Qwen 4 Coder, Max, and Flash: what exists

Named, not shipped. At Apsara on September 22 Alibaba announced four Qwen 4 tiers: Max as the flagship, Plus as the balanced multimodal tier, Flash for low-latency, high-volume work, and a 27B open-weights model for local deployment. None has public specs, pricing, or a launch date, so it's a statement of direction, not a product; what each tier is likely to be is broken down in Qwen 4 Max vs Plus vs Flash vs 27B. Flash-Next remains the early look at the Flash tier's architecture, and the 27B is the successor to the current Qwen3.8-27B; what's known about it is in Qwen 4 27B: what's confirmed and how to prepare. Searches for "Qwen 4 Coder" keep rising and there is still no such model; the coding-focused Qwens are the Qwen3-Coder line.

Open weights are the other open question. The 3.8 generation shipped mixed: the 27B under Apache 2.0, the Max checkpoint under a custom license, and Flash-Next under qwen-community-1.0. Expect Qwen 4 to be similarly mixed, and read the license on whichever tier you plan to build on.

How to prepare

Three things you can do now that pay off the day Qwen 4 lands.

Keep your serving behind an OpenAI-compatible interface. Every recent Qwen transition has been a model-string change for teams set up this way and a migration for teams that weren't. Our Qwen 3.8 production guide covers the pattern; the current Qwen line is live on Yotta AI Gateway alongside DeepSeek V4, GLM 5.3, and Kimi K3, which makes the eventual Qwen 4 A/B a routing rule.

Test the architecture early. Flash-Next is downloadable today, and running it against your workload is the only way to learn how the n-gram layer behaves on your data before the flagship version of it arrives.

Plan hardware for the memory rule, not the activation count. Whatever Qwen 4's tiers turn out to be, a low active-parameter figure will make the headlines and the total parameter count will decide your GPU bill. The open Flash-class comparison shows how differently three lean-activation models size out in practice. The Qwen 3.8-Flash-Next hardware requirements post has the verified GPU configs for the preview.

Frequently asked questions

When is Qwen 4 coming out?

Part of it already has. Qwen4-Max went into preview on November 11, 2026, on Alibaba's Token Plan, and the preview runs through November 25. Alibaba hasn't given a date for the full release or for Plus, Flash, and the 27B. Its launch post says only "Qwen4 is coming."

Is Qwen 4 out?

Partly. Qwen4-Max-Preview has been available since November 11, 2026, through Alibaba's Token Plan subscription and in Alibaba's own apps. The general release, the API on pay-as-you-go pricing, and the other three tiers are not out yet.

How do I try Qwen4-Max-Preview?

Subscribe to Alibaba's Token Plan and set the model ID to qwen4-max-preview in a supported coding tool, or use the early access in QwenWork, Qoder, the Qwen app, or QwenStudio. Pay-as-you-go API access is not supported during the preview, which ends November 25.

Are there Qwen 4 benchmarks?

Not from Alibaba. The preview launched without published scores, so any Qwen4-Max benchmark numbers circulating now are third-party tests of the preview or leaks, and worth treating with care until Alibaba or an independent lab publishes results.

Is Qwen 3.8-Flash-Next the same as Qwen 4?

No. Alibaba calls it an experimental preview of the Qwen 4 architecture, released early so developers can prepare. Qwen 4 itself is a separate model family, with only Qwen4-Max in preview so far.

Will Qwen 4 be open source?

Unknown. The 3.8 generation shipped mixed licenses across tiers, Apache 2.0 for the 27B, custom for the Max checkpoint, a community license for Flash-Next. Expect a similar mix and read the license per tier.

Is there a Qwen 4 Coder?

Not yet. The only Qwen 4 model available is the Qwen4-Max preview, which Alibaba describes as built for coding and agent work. The current coding-focused models are the Qwen3-Coder line.

What will Qwen 4's architecture be?

Based on Alibaba's own description of Flash-Next: a lean-activation sparse MoE with a large n-gram memory component, native multimodality, and 1M-scale context. How the flagship scales those choices is unannounced.

How big will Qwen 4 be?

Unknown. Qwen 3.8-Max was 2.4 trillion parameters; Flash-Next is 125B plus a 51B n-gram table. Any Qwen 4 parameter count you see is speculation until Alibaba publishes one.

Is Qwen 4 on Yotta?

During the preview, Alibaba is offering Qwen4-Max only through its own Token Plan. The current Qwen line, including 3.8-Max and 3.8-27B, is live on Yotta AI Gateway, and this post will be updated when Qwen 4 ships and again if it joins the catalog.

Bottom line

Qwen 4 has started to arrive. Qwen4-Max is in preview on Alibaba's Token Plan through November 25, with a 1M-token context and no published benchmarks or API price yet. The full release, the other three tiers, and any open weights still have no date, and the rollout is being staged the way 3.8's was.

The teams that did well on the last Qwen wave weren't the ones who guessed the date. They were the ones whose serving was ready for a model-string change and who had already tested the previous generation on their own workload. The current Qwen line on Yotta AI Gateway is one API key away, and Flash-Next is one download away. That's the preparation.