Close Menu
    Trending
    AI News

    What We Know So Far

    aitoday7By No Comments14 Mins Read
    What We Know So Far

    Qwen 4 is not out yet. As of 5 November 2026 there is no Qwen 4 model card, download, API name or price. The only thing Alibaba has put in writing is that Qwen 4 “is currently in training”. The closest thing to a Qwen 4 release date is one sentence from a conference audience. An attendee at Alibaba Cloud’s Apsara Conference in Amsterdam on 29 October reported the company’s CTO saying “Qwen4 next week”, which would mean 2 to 8 November.

    That week is now half gone, and the claims around it have multiplied.

    There are four Qwen 4 model names on two stage slides. There are Qwen 4 settings already merged into Alibaba’s own coding tool. And there is an open-weights question that several sites have answered with the wrong model’s news. This guide separates what Alibaba said from what was reported, grades every Qwen 4 rumor, and explains what to use, and what to watch, until Qwen 4 actually ships. We will update it the day it does.

    The Key Takeaways

    • Not released: Qwen’s Hugging Face page has no Qwen 4 repository. Its newest upload is still Qwen-Image-2.1-Turbo from 9 October 2026.
    • Official status: Alibaba’s 22 September press release says Qwen 4 is “currently in training”, with Qwen 4.5 and Qwen 5 planned at 5 to 10 trillion parameters. No date.
    • “Next week”: one attendee in Amsterdam quoted CTO Feifei Li saying “Qwen4 next week” on 29 October. Alibaba’s own write-up of that event does not mention Qwen 4.
    • The lineup: two Qwen slides read “Qwen4 Series Coming Soon” over Qwen4-Max, Qwen4-Flash, Qwen4-Plus and Qwen4-27B.
    • The odds: Manifold traders put a release before December at 93.5% (5 November), on a thin market of 60 traders.

    Is Qwen 4 Out?

    No. Qwen 4 has not been released as of 5 November 2026, 14:00 UTC. The Qwen organization on Hugging Face, where Alibaba publishes its open Qwen models, has no repository with Qwen 4 in its name. Its newest upload is Qwen-Image-2.1-Turbo, an image model added on 9 October. The official Qwen account on X, @Alibaba_Qwen, has not posted a Qwen 4 announcement, and Qwen Cloud, Alibaba’s paid API, lists no Qwen 4 model or price.

    The current Qwen generation is Qwen 3.8. Qwen3.8-Max, the flagship, launched on 3 August 2026, and Alibaba published its weights as Qwen3.8-2.4T-A95B shortly after. Our Qwen 3.8 review covers what that model can do and how it compares with the Western flagships.

    One caution before the details. At least one model database listed a “Qwen 4” release in June 2026, and copies of that entry still circulate. No such model exists on any Qwen repository, so any page telling you Qwen 4 came out over the summer is repeating a bad record.

    Qwen 4 Release Date: When Is It Coming Out?

    Alibaba has not announced a Qwen 4 release date. The best available signal is an attendee’s report of “next week” from 29 October, which points to 2 to 8 November 2026. Code changes in Qwen’s own tools back it up. The evidence comes in three layers, and they deserve different amounts of trust.

    What Alibaba Has Said in Writing

    Alibaba’s only written statement on Qwen 4 is one clause. At the Apsara Conference in Hangzhou on 22 September 2026, the company’s roadmap release said its next-generation model, Qwen 4, “is currently in training”. The same release announced a roadmap for Qwen 4.5 and Qwen 5, “projected to scale up to 5 to 10 trillion parameters”. It gave no date, no sizes, no tier names and no licence for Qwen 4 itself.

    The 10 trillion figure is not Qwen 4. It belongs to the two generations after it, and Alibaba gave no dates for those either.

    The “Next Week” Report From Amsterdam

    The “next week” date comes from a single attendee, not from Alibaba. On 29 October 2026 Alibaba Cloud held its first international Apsara Conference at the Beurs van Berlage in Amsterdam. Alibaba Cloud’s own account promoted the event the day before:

    That morning, attendee John T Davies posted the line that started the countdown, attributing it to Alibaba Cloud CTO Feifei Li:

    Alibaba’s written account of the same day is a press release about its first cloud region in the Netherlands. It quotes Feifei Li on European demand and lists partners such as JetBrains, which it says gets “early access to frontier Qwen models” for its Junie coding assistant. It does not mention Qwen 4. So the date rests on one person’s note of a remark made on stage, with no Alibaba signature under it.

    What the Prediction Markets Say

    Prediction markets expect Qwen 4 within weeks, not months. On Manifold, a market on the Qwen 4 release date gave 93.5% to a release before December 2026 when we checked on 5 November. A second market gave 83% to Alibaba publishing open weights for a Qwen 4 model on Hugging Face by 25 November. Both markets are small, with 60 and 6 traders, so treat them as a crowd’s mood rather than inside knowledge.

    Qwen 4 Models: Max, Flash, Plus and 27B

    Alibaba has shown four Qwen 4 model names, all on stage slides and none in writing. A slide shown in Hangzhou on 22 September read “Qwen4 Series Coming Soon” in Chinese and English over three boxes: Qwen4-Max, Qwen4-Flash & Qwen4-Plus, and Qwen4-27B. In Amsterdam on 29 October, Davies photographed the same slide during a talk he credits to Xudong Guo, a research scientist on the Qwen team:

    Alibaba has published no sizes, prices or positioning for the four tiers. The table reads each name against what the same tier is in the Qwen 3.8 generation, which is the only precedent available.

    A fifth name keeps appearing in searches: Qwen 4 Coder. Alibaba has not shown a Coder tier. The identifier qwen4-coder-plus does appear in Qwen’s own code, which is the next piece of evidence.

    What Qwen’s Own Code Says About Qwen 4

    Qwen’s code shows a team preparing its tools for Qwen 4, but it does not show a release date or a finished spec sheet. The evidence sits in Qwen Code, Alibaba’s open- the week of the Amsterdam event

    The 1M-Token Setting

    On 28 October, Qwen Code merged pull request 15762, titled “size the qwen4 generation at 1M”. It tells the tool to treat every Qwen 4 identifier as a model with a 1,000,000-token context window instead of the 262,144-token default. The identifiers it tested include qwen4-max, qwen4-plus, qwen4.5-max, qwen4-coder-plus and qwen4-omni-120b-think. The author says no Qwen 4 entry existed yet in the public model catalog Qwen Code reads. The tests ran against a fake server, not a live Qwen 4 endpoint.

    The same day, a sister change sized a set of invitation-only previews, named like qwen-latest-series-invite-202610-m4, at the same 1M tokens. Nothing in either change says those previews are Qwen 4.

    The Reasoning-Effort Tiers

    On 31 October, a third change gave the whole qwen4 family and the invitation-only previews the same reasoning controls as the current Qwen models. In Qwen Code, Qwen 4 identifiers now get a thinking toggle with low, medium and xhigh effort levels, and xhigh as the default. Reasoning effort is how long a model works through a problem before it answers.

    What the Code Does Not Prove

    The Qwen Code changes prove that Qwen’s engineers expect Qwen 4 identifiers to reach their tools soon and want them to work on arrival. They do not prove a date, and the 1M figure is a tool setting, not a published model specification. The real context window will be whatever the first Qwen 4 model card says.

    The Qwen 4 Architecture You Can Already Download

    Qwen 4’s design is already public, inside a model you can download today. On 26 August 2026 the Qwen team released Qwen3.8-Flash-Next and wrote on GitHub that it “serves as an early preview of the architecture used in Qwen4”. The team added that it “plays the same role that Qwen3-Next played for Qwen3.5”. It said it was “again releasing the architectural changes early” so the community could examine them before the full Qwen 4 family is built on top.

    The Qwen3.8-Flash-Next model card explains the main bet. Instead of spending compute on every parameter, Qwen3.8-Flash-Next parks a large share of them in an n-gram lookup table. That table is cheap to read and can sit in ordinary memory. Only 6 billion parameters do work on each token. If Qwen 4 keeps that shape, its models should be cheaper to run than their total size suggests.

    Open-llama.cpp project, which runs models on ordinary computers including Macs, merged support for Flash-Next on 27 August under the internal name “qwen4exp”. That is engineering vocabulary for the architecture, not evidence of a Qwen 4 release

    Will Qwen 4 Be Open Source?

    Nobody knows yet, because Alibaba has not announced a Qwen 4 licence. The Qwen 3.8 precedent says to expect open weights for several tiers, but under three different licences, and the tier you care about decides which one applies.

    What the Qwen 3.8 Licences Say

    The Qwen 3.8 generation shipped open weights under three licences. Qwen3.8-27B uses Apache 2.0, the most permissive option, so you can use it commercially without conditions beyond keeping the notice.

    Qwen3.8-Flash-Next uses the Qwen Community License 1.0. Any business that sells model access through an API, or builds a coding or office assistant, needs a separate licence from Qwen for commercial use.

    Qwen3.8-Max shipped its weights as Qwen3.8-2.4T-A95B under its own licence. Products above 100 million monthly users or US$20 million in monthly revenue must display the model name. API and assistant businesses need a separate licence once they pass US$50 million in revenue over 12 months.

    So the common claim that Max models stay API-only is out of date. Alibaba’s model card for Qwen3.8-2.4T-A95B says Qwen 3.8 “brings a Qwen-Max-class model to open release” for the first time.

    Some Qwen 4 coverage cites a Reuters report that Alibaba wants a cut of revenue from big users of its next open model. That report, from August 2026, named Qwen3.8-Max, not Qwen 4, and said the rate was still under discussion. What actually shipped with Qwen3.8-Max is the licence above: name display and a separate licence for the largest API and assistant businesses. If you plan to build on a Qwen 4 model, read the licence on its model card on release day, because Qwen’s licences now differ by tier.

    Qwen 4 Rumors, Graded

    Most Qwen 4 rumors are either true but unofficial, or about a different model. Here is every claim in wide circulation, checked against Alibaba’s pages, Qwen’s repositories and the original posts.

    What Qwen’s Release History Says About the Qwen 4 Release Date

    Qwen’s release history says a surprise drop is likely and a slip is normal. The team usually ships without warning, with an X post and the weights already on Hugging Face. That habit is why the first sign of Qwen 4 will probably be a new repository, not a press release.

    The team’s own precedent for an architecture preview argues for patience. Qwen3-Next, which previewed the design behind Qwen 3.5, appeared on Hugging Face in September 2025. The Qwen 3.5 family followed in February 2026, about five months later. Flash-Next previewed Qwen 4 on 26 August 2026. The same gap would put Qwen 4 in early 2027, so a November release would mean the team moved twice as fast this time.

    The Qwen 3.8 generation shows the team can move that fast. Qwen3.8-Max launched on 3 August, its open weights followed later in August, and Flash-Next arrived on 26 August. Alibaba also says it ran Qwen3.8-Max through 33 automated self-improvement cycles over a month, raising its Artificial Analysis score from 40 to 45. That is Alibaba’s own figure.

    Three places will confirm Qwen 4 first: the Qwen organization on Hugging Face, the @Alibaba_Qwen account on X, and the Qwen blog at qwen.ai. If none of them shows Qwen 4 by Sunday, 8 November, the “next week” remark will have slipped. That would not be unusual for a launch date spoken on stage.

    Will Qwen 4 Run on a Mac?

    Qwen4-27B probably will, and Qwen4-Max almost certainly will not. Alibaba has published no Qwen 4 sizes beyond the 27B in that model’s name, so the current Qwen3.8-27B is the best yardstick. Its 4-bit build for Apple’s MLX framework is 16.05 GB of weights on Hugging Face. That fits on an Apple silicon Mac with 32 GB of unified memory, with room left for the system and a long conversation.

    If Qwen4-27B ships with open weights and a similar size, expect Mac-ready builds quickly. llama.cpp already supports the Qwen 4 architecture preview, and the MLX community has published 4-bit builds of Flash-Next. Our guide to the Mac Studio for local AI lists measured model sizes against memory configurations.

    Qwen4-Max is a different matter. Its predecessor’s open weights are 2.4 trillion parameters, which is data-center territory. On a Mac, you will reach Max-class Qwen models through an app or an API, not a download.

    What to Use While You Wait for Qwen 4

    Qwen 3.8 is the model to use until Qwen 4 ships, and it is not a stopgap. Qwen3.8-Max is available in Qwen Studio at chat.qwen.ai and costs $2 per million input tokens and $6 per million output tokens on Qwen Cloud. Qwen3.8-Flash costs $0.15 and $0.47 for the same amounts. Our Qwen pricing guide breaks down the free app and the API plans.

    For local use, Qwen3.8-27B is the Qwen model most Apple silicon Macs with 32 GB can run today, and Flash-Next lets you test the Qwen 4 architecture today if you have the memory for it. Our list of the best Chinese AI models shows how Qwen compares with Kimi, DeepSeek and GLM. The DeepSeek V5 release date guide tracks the other big Chinese launch everyone is waiting for.

    On a Mac, our Qwen app for Mac page compares Qwen’s official app with the alternatives. Fello AI already puts Qwen in the same window as ChatGPT, Claude and Gemini. New models usually reach it as soon as their API opens, so trying Qwen 4 should not mean another sign-up.

    The Bottom Line

    Qwen 4 is close, but it is not here, and its release date still rests on one remark nobody at Alibaba has repeated in writing. Everything else points the same way. Two Qwen slides say “coming soon”, Qwen’s own coding tool has been updated to handle Qwen 4 models, and prediction markets give a release before December better than nine-in-ten odds. A release this week is a fair bet, and a release this month a strong one. A slip into December would surprise no one who remembers the five months Qwen 3.5 took.

    Until a Qwen 4 model card appears, ignore any page quoting Qwen 4 benchmarks or a Qwen 4 licence. When Qwen 4 lands, the questions that matter are which tiers get open weights, under which licence, and whether Qwen4-27B fits on your Mac. We will answer all three here on release day.

    Is Qwen 4 out?

    No. As of 5 November 2026 there is no Qwen 4 model card, download, API name or price. Alibaba’s only written statement says Qwen 4 is “currently in training”, and the newest model on Qwen’s Hugging Face page is Qwen-Image-2.1-Turbo from 9 October.

    When is Qwen 4 coming out?

    Alibaba has not announced a date. An attendee at the Apsara Conference in Amsterdam on 29 October 2026 reported CTO Feifei Li saying “Qwen4 next week”, which would mean 2 to 8 November. Prediction markets put a release before December at about 93%.

    What are the Qwen 4 models?

    Qwen slides shown in Hangzhou and Amsterdam name four: Qwen4-Max, Qwen4-Flash, Qwen4-Plus and Qwen4-27B. Alibaba has published no sizes, prices or release dates for any of them.

    Alibaba has not announced a Qwen 4 licence. The Qwen 3.8 generation opened weights for its Max, Flash-Next and 27B models under three different licences, from Apache 2.0 for the 27B to custom terms for Max. Expect a similar mix.

    Will Qwen 4 run on a Mac?

    The 27B tier probably will. The current Qwen3.8-27B weighs 16.05 GB in a 4-bit MLX build, which fits a 32 GB Apple silicon Mac. Qwen4-Max is likely far too large to run locally, so expect to use it through an app or the API.

    Post Views: 6

    Share.

    Related Posts

    Add A Comment
    Leave A Reply