Qwen4-Max Preview: Alibaba opens Qwen4, no benchmarks, two weeks only
Qwen4-Max Preview is live in Qwen Studio, the Qwen app, Qoder and via the Token Plan API. 1M context, preview until 25 Nov, 8% night rate, no benchmarks yet.
👁 112 reads
On 11 November 2026 at 02:01 UTC the Qwen team posted “Qwen4 is coming” and opened a preview of the first model of the new generation. Qwen4-Max Preview is available from today in the Qwen Studio chat, in the Qwen app, in QwenWork and Qoder, and through the API on Alibaba Cloud’s Token Plan subscription. The preview is short: Alibaba will retire it after 25 November. There are no benchmarks, no per-token price and no OpenRouter listing yet, so today it cannot be measured the way I measure other models.
Everything about the new family is in my Qwen4 guide.
What Alibaba announced
The announcement is short and has no numbers. Qwen4 promises “deeper knowledge”, “richer world understanding” and “stronger coding and agentic capabilities”, built for real-world tasks: from complex coding and knowledge work to front-end builds and professional software workflows. The model description in Qwen Studio adds reasoning, advanced mathematics, long multi-step tasks with tools and “visually polished, production-ready interfaces”.
- API model ID:
qwen4-max-preview(the aliasqwen4.0-max-previewalso works). - Context window: 1,000,000 tokens (Qwen Studio model list and Alibaba documentation).
- Input: text, images and video; in the chat also documents and audio. Output: text.
- Thinking: switches on automatically in the chat. In the API, max_tokens for this model counts thinking and the answer together.
- Benchmarks: none published by Alibaba. Qwen4 is not yet on the Artificial Analysis leaderboard.
- Weights: Alibaba says nothing about open weights. The model is available only through Alibaba services.
Where it runs and for how long
| Where | How | Account |
|---|---|---|
| Qwen Studio (chat.qwen.ai) | “Qwen4-Max-Preview” in the model list, with thinking, search, web dev, deep research and slides modes | yes, not available to guests |
| Qwen app | early access | yes |
| QwenWork, Qoder | early access in Alibaba’s work and coding tools | yes |
| API: Alibaba Cloud Model Studio | Token Plan subscription only, Singapore region, no pay-as-you-go | yes, Alibaba Cloud account |
| OpenRouter | not listed (11 November, 08:20 UTC) | - |
The Token Plan preview runs from 11 November 10:00 to 25 November 24:00 Beijing time (UTC+8), that is from 02:00 UTC on 11 November to 16:00 UTC on 25 November. During this window Alibaba charges fewer credits: as low as 20% of the standard rate, and during Beijing night hours (22:00-08:00 UTC+8, i.e. 14:00-00:00 UTC) only 8%.
What it costs
There is no price per million tokens because the model is not offered on pay-as-you-go billing. You pay for a Token Plan and get a pool of credits shared by all models in the plan.
| Plan | Monthly price | New users until 30 Nov | Credits |
|---|---|---|---|
| Lite | $6 | $3 | 20,000 |
| Essential | $10 | $5 | 45,000 |
| Standard | $18 | - | 80,000 |
| Pro | $43 | - | 200,000 |
Token Plan Personal, Alibaba Cloud, read on 11 November 2026. Extra bundle: $15 for 20,000 credits.
How many credits one Qwen4-Max request uses, Alibaba does not say; the console shows it only after use. For comparison, its predecessor Qwen3.8 Max costs $2 per million input tokens and $6 per million output tokens.
Why there is no test from me
I run every new model through the same set of tasks via OpenRouter so that results are comparable. Qwen4-Max Preview is only available after creating an Alibaba account and buying a subscription, and the chat cannot be used without signing in. I will test it once it reaches the API on normal terms.
What it means
Alibaba is doing with Qwen4 what Google did with Gemini 4: show the strongest model first, with limited access and without a results table. A short preview with an 8% night rate looks like collecting feedback from developers before the full launch. The smaller models of the family have no dates yet. On the Artificial Analysis index the current leaders are Claude Fable 5.2 (65), Gemini 4 Carbon (60) and GPT-6 Astra (57); its predecessor Qwen3.8 Max scored 45 in September, so Qwen4-Max would need a big jump to catch up.
Sources
- Qwen on X, post of 11 November 2026.
- Token Plan (Personal Edition), Alibaba Cloud Model Studio: prices, preview window, discounts.
- Model Studio model list.
- Qwen Code in Model Studio: model ID and output limit.
- Qwen Studio: chat model list (11 November 2026).
- OpenRouter: no Qwen4 listing (11 November 2026).
More on this from QUANTSLANT
ai tools·ACE Music: AI song generator, pricing, free plan and my testACE Music in one place: the #1 v1 model, plans and prices, the free tier, my lyric test across genres with audio, covers, sheet music, ACE-Step and Suno.