artificialwatch
model radar / alibaba / qwen-4.0-max

Qwen 4.0 Max

Alibaba · the new numbered flagship from China's top lab
live nov 16 · $1.50 / $4.50 107 days after 3.8 max 1m context · text, image, video in
Watch this model free →Free alert about 15 min after it answers on a public API · members the second it does · no card

Qwen 4 Max is out. Alibaba released it on November 16 at 02:01 UTC, with the API live on its QwenCloud platform as qwen4-max: $1.50 / $4.50 per million input and output tokens, $0.075 for cached input, and a 1M-token context with up to 393K output tokens. It takes text, image and video and returns text. Alibaba calls it a new MoE design and says it beats Qwen 3.8 Max with up to about half the tokens. The Max weights are not public; Alibaba says open weights for smaller Qwen 4 models will follow. It came 107 days after Qwen 3.8 Max (August 1) and five days after the two-week Qwen4-Max-Preview reached Qwen Chat (November 11). ArtificialWatch’s radar saw qwen4-max 50 seconds before Alibaba’s announcement.

✳ the wire

Leaks & news, newest first.

Nov 16 · official

Qwen 4 Max is out: $1.50 / $4.50, 1M context, on Alibaba’s API

At 02:01 UTC @Alibaba_Qwen announced Qwen4-Max, with the API live on QwenCloud as qwen4-max. QwenCloud’s model page lists $1.50 per million input tokens, $4.50 per million output, $0.075 for implicit-cache reads, a 1M-token context with up to 991K input and 393K output, and text, image and video input. It is also on OpenRouter. Alibaba’s own chart puts it ahead of Claude Opus 5.5 on 6 of 16 benchmarks, among them DeepSWE v1.1 (75.7 vs 74.2), and behind on 10, including Terminal-Bench 4.0 (48.5 vs 59.6) and Humanity’s Last Exam (51.7 vs 59.8). Those are Alibaba’s figures, not independent tests. ArtificialWatch’s radar first saw qwen4-max at 02:01:05 UTC, 50 seconds before the post, and fired at 02:02:07 UTC. Our write-up. via @Alibaba_Qwen on X · QwenCloud · OpenRouter

Nov 11 · official

Qwen4-Max-Preview goes live in Qwen Chat, as a two-week preview

Qwen Chat’s public model list carries qwen4.0-max-preview (“Qwen4-Max-Preview”, with a 1M-token context), and Alibaba Cloud Model Studio offers it through its Token Plan from 02:00 UTC Nov 11 to 16:00 UTC Nov 25, when the preview is retired. There is no pay-as-you-go price and no OpenRouter id, and Alibaba had not announced it by 03:53 UTC. ArtificialWatch’s radar caught it on Qwen Chat’s list and sent an early-version alert at 02:11 UTC; the main Qwen 4 Max alert stays armed for the final model. Our write-up. via chat.qwen.ai · Alibaba Cloud Model Studio

Nov 8 · status

The last day of the CTO’s “next week”, and still no Qwen 4

Nov 8 is the last day of the Nov 2 to 8 window that the CTO’s Oct 29 “Qwen4 next week” implies (one attendee’s account), and at 02:20 UTC no qwen4 id is on Hugging Face (Qwen’s newest repo is still Qwen-Image-2.1-Turbo, from Oct 9), OpenRouter or models.dev. Qwen Chat, where Qwen ships first, lists Qwen3.7-Plus, Qwen3.8-Max and Qwen3.8-Omni-Flash. @Alibaba_Qwen has posted nothing since the Nov 6 teaser, and its pinned Aug 26 post still introduces Qwen3.8-Flash as “an early preview of the Qwen4 architecture”. Qwen Code shipped v0.27.0 on Nov 7 with no qwen4 change, and has merged none since Oct 31. Polymarket’s “next Qwen Max (3.9+)” market last traded at 61% for a release by Nov 9, 77% by Nov 11 and 94% by Nov 30. It has been 99 days since Qwen 3.8 Max. via Hugging Face · OpenRouter · models.dev · chat.qwen.ai · @Alibaba_Qwen on X · QwenLM/qwen-code · Polymarket

Nov 7 · status

A day after the teaser, still no Qwen 4 in any catalog we sweep

On Nov 7, day six of the Nov 2 to 8 window the CTO’s “next week” implies, no qwen4 id is on Hugging Face (Qwen’s newest repo is still Qwen-Image-2.1-Turbo, from Oct 9), OpenRouter or models.dev, and Qwen Code has merged no new qwen4 change since Oct 31. The open-weight Qwen4-27B that @Mr_Salio said on Nov 6 was “dropping today” has not appeared. Polymarket’s “next Qwen Max (3.9+)” market, which a Qwen 4 Max would settle and a 27B would not, last traded at 61% for a release by Nov 9, 81% by Nov 13 and 92% by Nov 30. It has been 98 days since Qwen 3.8 Max. via Hugging Face · OpenRouter · models.dev · QwenLM/qwen-code · Polymarket

Nov 6 · official

Qwen’s account posts a teaser: “The becoming begins.”

At 12:00 UTC @Alibaba_Qwen posted “The becoming begins.” with an image of a flock of birds over the sea at dusk and the line 北冥有鱼 其名为_ (“In the northern sea there is a fish; its name is _”), the opening of the Zhuangzi, where the fish Kun becomes the giant bird Peng. The post names no model and no date, and a widely shared “Qwen 4” graphic did not come from it. @Mr_Salio, @LuminaBench and TestingCatalog read it as Qwen 4, with Max, Plus, Flash and an open-weight 27B expected and the 27B first (unconfirmed). Our write-up. via @Alibaba_Qwen on X

Nov 4 · status

Halfway through “next week”, and no Qwen 4 anywhere

Read literally, the CTO’s Oct 29 “next week” means Nov 2 to 8. On Nov 4 no qwen4 id is on Hugging Face (Qwen’s newest repo is Qwen-Image-2.1-Turbo, from Oct 9), OpenRouter or models.dev, and Alibaba’s Qwen Code repo has merged no new qwen4 change since Oct 31. It has been 95 days since Qwen 3.8 Max. Polymarket’s “next Qwen Max (3.9+)” market, which a Qwen 4 Max would settle, last traded at 94% by Dec 31. via Hugging Face · OpenRouter · models.dev · QwenLM/qwen-code · Polymarket

Oct 31 · code

Invite-only test ids, grouped with Qwen 4

Alibaba’s open-source Qwen Code repo groups invite-only qwen-latest-series-invite aliases with the qwen4 ids for reasoning effort (PR #15987), and has sized them at 1M tokens since Oct 28. @Mr_Salio reads them as invite-only Qwen 4 testing ahead of a public release next week. Unconfirmed: the alias names no generation. via QwenLM/qwen-code on GitHub · @Mr_Salio on X

Oct 29 · signal

Alibaba Cloud’s CTO: “Qwen4 next week”

At Alibaba Cloud’s first international Apsara Conference, in Amsterdam, attendee @jtdavies quoted CTO Feifei Li on stage: “Qwen4 next week”, then posted “Qwen4, coming soon” from a talk by Qwen research scientist Xudong Guo. One attendee’s account: Alibaba’s own posts from the event cover its new Netherlands cloud region, not Qwen 4. via @jtdavies on X

Oct 28 · code

Qwen Code sizes every qwen4 id at 1M tokens

Alibaba’s coding agent merged a rule giving any model id that starts with qwen4 a 1,000,000-token context as a “forward default”, with tests naming qwen4, qwen4-max, qwen4-plus and qwen4-coder-plus. It is the tool’s default, not a published spec. Our write-up. via QwenLM/qwen-code on GitHub

Oct 27 · signal

A second LMArena checkpoint, “Kerria”

@Mr_Salio reports another Qwen 4 checkpoint on LMArena, under the codename Kerria. Unconfirmed, and Alibaba has said nothing. via @Mr_Salio on X

Oct 26 · signal

“qwen4-exp” turns up in a code list

@Mr_Salio posted a screenshot of a code list naming qwen4-exp next to qwen3.8-flash-next, without saying where it comes from, and expects Qwen 4 this week or next. Unconfirmed, and Alibaba has said nothing. Our hunt fires on qwen4-exp. via @Mr_Salio on X

Oct 22 · signal

Testers report Qwen 4 on LMArena as “Kalmia”

@LuminaBench and @Mr_Salio say Qwen 4 is being tested on LMArena under the codename Kalmia, possibly in more than one version. Unconfirmed: Alibaba has said nothing, and no Qwen 4 id is in OpenRouter or models.dev. via @LuminaBench · @Mr_Salio on X

Sep 22 · official

Alibaba: Qwen 4 is in training

At the Apsara Conference, Alibaba said its next-generation model, Qwen 4, “is currently in training,” and laid out a roadmap for Qwen 4.5 and Qwen 5 “projected to scale up to 5 to 10 trillion parameters.” No release date, pricing or benchmarks. via Alibaba Cloud

Aug 12 · weights

The 2.4T base goes open-weight

Qwen published Qwen3.8-2.4T-A95B on Hugging Face, the base of Qwen3.8-Max. Per the model card, the hosted Max adds vision input, non-thinking support, 1M context by default and built-in tools. via Hugging Face

Aug 3 · launch

Qwen3.8-Max launches: 2.4T MoE, 1M context

Alibaba announced Qwen3.8-Max, which it calls its largest flagship to date: a 2.4T-parameter sparse MoE with 95B active, a 1M-token context and API access through Model Studio, with weights “scheduled for release next week.” Our sweep had already seen it served as the flagship on chat.qwen.ai on Aug 1. via Alibaba Cloud

Jul 7 · signal

The market keeps Alibaba on top — 4.0 Max is the model it's pricing

At 96%, the top-CN-lab market is effectively settled sentiment. The open question is when the 3.7 line rolls over to 4.0 — and whether an open-weight Qwen 4 ships beside the Max.

May 20 · launch

Qwen 3.7 Max: 1M context, and it beats Opus 4.6 Max on the agentic benches

Announced at the Alibaba Cloud Summit in Hangzhou: 1M-token context, a 56.6 on the Artificial Analysis index, wins over Claude Opus 4.6 Max on Terminal-Bench 2.0, SWE-Bench Pro and MCP-Atlas, and a claimed 35-hour autonomous run across 1,000+ tool calls — at $2.50/$7.50 per 1M tokens. via Alibaba · Digital Applied

May 19 · detail

The API went live a day before the keynote

Model Studio served 3.7 Max on May 19 — a day ahead of the announcement — with OpenRouter and Together AI cross-listings on day zero. Alibaba ships API-first, which is exactly the surface we sweep. via Alibaba Cloud Model Studio

ongoing · status

Qwen 4.0: in training, no date — cadence is still the tell

Alibaba’s only official word is that Qwen 4 is in training (Sep 22). 3.6 → 3.7 took a quarter and 3.7 → 3.8 about ten weeks, and Alibaba hasn't let a flagship sit two full quarters since the Qwen 3 era. via Alibaba Cloud

✳ from alibaba

The official word.

"Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window."

@Alibaba_Qwen on X · Nov 16, 2026 — API live on QwenCloud

Qwen 3.8 MaxAlibaba’s flagship from August until Qwen 4 Max shipped Nov 16 — 2.4T MoE, 95B active; base weights open since Aug 12.
Qwen 4.0 MaxLive since Nov 16 on QwenCloud: $1.50 / $4.50 per 1M tokens, 1M context.
Qwen 4 (open)Alibaba says open weights of smaller Qwen 4 models will follow; none are out yet.
✳ where to use it

Listed by 16 providers, exact ids below.

16 providers list Qwen 4 Max in the models.dev catalog (as of 2026-11-20), under the model ids shown below.

OpenRouterqwen/qwen4-max
Vercel AI Gatewayalibaba/qwen4-max
Charm Hyperqwen4-max
CrossModelqwen/qwen4-max
Deep InfraQwen/Qwen4-Max
DevPass (LLM Gateway)qwen4-max
Eden AIqwen/qwen4-max
EmpirioLabs AIqwen4-max
Kenariqwen4-max:free
Kilo Gatewayqwen/qwen4-max
LLM Gatewayalibaba/qwen4-max · scx-ai-gp/qwen4-max
NanoGPTqwen/qwen4-max · qwen/qwen4-max:thinking

Also listed by Novita AI, Ofox, OpenCode Go, Venice AI.

Listings cover the model family, so a provider may list a sibling tier under the same name. Check the id before you build on it.

Hear the next Qwen the second it answers.

Qwen 4 Max is out, and Qwen 4 Plus, Flash, the open-weight 27B and Qwen 4.5 are next. We sweep chat.qwen.ai's public model list and the open catalogs every 60 seconds; that sweep saw qwen4-max 50 seconds before Alibaba announced it. SMS and a phone call the moment it lands on paid tiers; free email + Chrome push about 15 minutes later.

On the board. You'll hear the second the next Qwen answers.