Qwen 4 Release Date: Max, Plus, Flash, 27B & What We Know

Qwen 4 is teased but not released. On November 6, 2026 the official Qwen account posted "The becoming begins", and since then everyone has been asking one thing: is it today? As of November 9 there is still no model card, download or API price.

This page tracks the Qwen 4 release date, the four Qwen 4 tiers, the leaked benchmark scores and what you can run now. We update it on the same URL the moment Qwen 4 ships.

Qwen 4 release date: what Alibaba has said

The "next week" window from Apsara (November 2–8) has passed with nothing new. That does not mean a delay: Qwen often ships late at night Beijing time and in stages. A Qwen meetup in Tokyo on November 11 is another date people watch.

The four Qwen 4 models

• Qwen4-Max: the flagship for hard reasoning and agent work, API only.
• Qwen4-Plus: the balanced multimodal tier.
• Qwen4-Flash: fast and cheap, for high-volume work.
• Qwen4-27B: the open-weight model, sized to run on one good GPU.

Leaks from the Qwen Code repo also name qwen4-coder-plus and qwen4-omni-120b, so the family may grow after launch.

Qwen 4 benchmarks: the leaked numbers

An AIbase report on November 9 says Qwen4-Flash scored 49.9 on one agent benchmark, above Opus 5.5 at xHigh effort (47.2), and that Qwen4-27B scored 44.6, close to Sonnet 5.5 (45). The same report puts the cost of one task at about $0.40 on Qwen4-Flash and $0.16 on Qwen4-27B, against $2.60 on Opus 5.5.

If those numbers hold, the story is price: near-frontier work for a fraction of the cost, plus an open 27B model you can host yourself.

Qwen 4 architecture: what the preview shows

Alibaba called Qwen3.8-Flash-Next "an experimental preview of the architecture that will underpin Qwen4". It is a sparse mixture-of-experts model with 125B total and 6B active parameters, a 51B n-gram memory table that swaps compute for lookup, and native multimodal input. Expect the new models to build on the same design.

Will Qwen 4 27B run on my PC?

Probably, with the right card. Today's Qwen3.8-27B is a 27B dense model that fits in about 17 GB at 4-bit. If Qwen4-27B keeps the same size, a 24 GB GPU or a 32 GB Mac should run it at 4-bit. Kernel code for Qwen4-27B dense has already landed in hardware repos.

Qwen 4 vs GPT, Claude and Gemini: who it is for

Most teams will not pick one model. The likely split: Qwen4-Max for hard reasoning when price matters, Qwen4-Flash for bulk work like tagging, extraction and agent sub-steps, and Qwen4-27B for anything that must stay on your own hardware. Claude Opus 5.5 and GPT-6.1 Sol stay the safe picks for long coding sessions until independent tests say otherwise.

Watch three things in the first week: how the open model does on real repos rather than one benchmark, whether the API is cheap outside China, and how fast tools like Ollama, vLLM and LM Studio add support. Past Alibaba launches got local builds within a day.

Qwen 4 Max Preview is out (Nov 11, 2026)

The first Qwen 4 model is here. Alibaba opened Qwen4-Max-Preview on November 11, 2026, as the flagship of the Qwen 4 family. The model ID is qwen4-max-preview.

• What it is: a multimodal model that reads text, images and video, thinks by default, and handles coding, office work, front-end design and agent tasks.
• Specs: 1M-token context and up to 393,216 output tokens, shared by thinking and the answer. Thinking effort can be low, medium or xhigh (default).
• How to get it: free in the Qwen app (pick Qwen4-Max-Preview in the model menu), or by API through Alibaba's Token Plan only; no pay-as-you-go price yet.
• Token Plan: Lite $6 a month ($3 for new users), Essential $10, Standard $18, Pro $43. Plans are sold by credits and concurrent agents.
• Short window: the preview runs to 24:00 on November 25, 2026 (UTC+8), then this version is retired.
• Still missing: no benchmark table, no model card, no open weights. Qwen 4 Plus, Flash and 27B are announced but not out.

If you want to test the Qwen 4 flagship, the Qwen app is the fastest way in. For production, keep your current model until Alibaba ships the full release and a per-token price.

Use top models in Sumus while Qwen 4 is on the way

Yes, you can stop waiting and get work done today. In Sumus, GPT, Claude and Gemini work as AI employees in one team chat, so you can run the same task on several models and keep the best answer. When the new Qwen lands, you compare it against the models you already use, in the same thread.

What we're watching

• A model card on Hugging Face or a Qwen blog post.
• Qwen4-27B weights and the license (Apache 2.0 or not).
• API names and prices on Qwen Cloud.