Qwen4-Max Preview: Cloud-Only at 80% Off, Local Users Must Wait for 27B Open Weights

Overview

On November 11, Alibaba's Qwen team released a preview version of Qwen4-Max exclusively through the Alibaba Cloud Token Plan subscription service. The preview is a cloud-only API offering — no model weights are available for download — and the company has not disclosed parameter counts, benchmark scores, or a formal release date. The launch is framed as an early-access program to gather real-world feedback on coding, agentic, and complex reasoning tasks.

What Qwen4-Max Preview Is

The official announcement highlights improvements in "knowledge understanding, world perception, coding, and agentic capabilities," citing internal testing results on complex programming, professional office work, frontend design, and professional software operation. First-party products including Qwen Office, Qoder, the Qwen App, and QwenStudio have already integrated the model. A screenshot from Qoder CN shows Qwen4-Max-Preview selectable in the model dropdown, confirming application-layer integration.

Qoder CN test: model dropdown shows Qwen4-Max-Preview (red box)

Table A: Six things to know about Qwen4-Max Preview

Three key omissions are noted: no specific parameter count, no benchmark results, and no formal release timeline. This suggests a strategy of validating real-task performance before publishing spec-sheet numbers.

Pricing Details

The Token Plan Double-11 promotion offers two personal tiers:

Lite : ¥19.5/month (regular ¥39), includes 20,000 Credits/month, 1–2 parallel agents, suitable for trial use.

Essential : ¥39.5/month (regular ¥79), includes 45,000 Credits/month, 2–3 parallel agents, adds web search and knowledge retrieval Harness tools, suited for advanced practical work.

Qwen4 invocation is discounted to 20% of standard price for a limited time; new users receive an additional 50% off (effectively 10% of list price). Monthly Credits quotas are ~78% higher than the previous plan (per Alibaba Cloud's November 9 plan adjustment notice), and existing subscribers are covered. Over 20 flagship models are bundled under a single subscription.

Table B: Available experience tiers under Token Plan 11-11 promotion

Local Deployment Constraints

Qwen4-Max Preview is strictly a cloud API; there are no weights to download, so it cannot run on local hardware (e.g., Mac/Apple Silicon). The Qwen4 family was introduced at Apsara 2026 with four named tiers: Max, Plus, Flash, and 27B. Only Max Preview is available now; the other three — including the 27B open-weight tier most relevant for local inference — have not been released.

Following the Qwen3.8 cadence (flagship API first, flagship weights ~10 days later, smaller models afterward), the 27B open-weight release is expected only after Max goes GA. The previous generation's Qwen3.8-27B (dense 27B, Apache 2.0, native 262K context extendable to 1M via YaRN, ~17 GB at Q4 quantization) fits on 24 GB VRAM and remains the only tier that is both commercially permissive and locally runnable. It is currently the "sweet spot" for 24 GB+ Mac users.

Table C: Local perspective — runnable Qwen is still 3.8 generation

Architecture Speculation (Not Official)

Based on the Qwen3.8-Flash-Next preview, Qwen4 may adopt an extremely sparse design: 125B main model + 51B n-gram embedding table, activating ~6B parameters per token, with native 262K context extendable to 1M via YaRN. The article explicitly labels this as extrapolation from the prior architecture, not confirmed fact, because Alibaba has released zero Qwen4 parameters or benchmarks.

Decision Guide

A decision matrix (Table D in the source) categorizes readers:

Coding/Agent early adopters : Spend ¥19.5 for the 20% discount to test real-task feel; preview exists for early feedback.

Local-only open-model users : Max is irrelevant; watch for Qwen4-27B open weights (estimated), meanwhile Qwen3.8-27B stays the sweet spot.

Production/SLA-dependent teams : Wait for GA; preview lacks SLA and published specs.

Observers : No urgency; decide after formal launch.

Table D: Decision matrix — should you pay ¥19.5 now?

Author's Recommendations

If you want to experiment with new coding/agent capabilities, the ¥19.5 entry is low-risk for hands-on validation.

If you rely on local Mac inference, ignore Max; track Qwen4-27B open-weight release and keep using Qwen3.8-27B.

If you need production stability and SLA, wait for the official release.

If you're just watching, there's no need to pay now.

Bottom line: Qwen4 is genuinely arriving, but today you can only touch a cloud preview at a discount; the tier that fits in your Mac still requires more waiting.

Data Sources

Release & pricing: Qwen official WeChat post 2026-11-11 10:44; Qwen AI platform activity page (tongyi.aliyun.com); IT Home / NetEase / Tencent News 11-11 coverage.

20% / 10% discounts, ~78% Credits increase: Official activity page and Alibaba Cloud 11-9 Token Plan adjustment announcement (via NetEase repost).

Qwen3.8 specs (3.8-Max 2.4T MoE / 95B active / 1M / $2·$6 per 1M; 3.8-27B dense 27B / Apache 2.0 / 262K→1M / Q4 ~17 GB): Qwen official model cards and multiple hardware reviews (since 2026-08).

Qwen4 four-tier naming, architecture inheritance from 3.8-Flash-Next: Apsara 2026 stage and Qwen official architecture blog. ⚠️ Qwen4 specific parameters / benchmarks / release date not officially disclosed; architecture discussion is estimation, not fact.

No benchmark numbers for Qwen4-Max were provided by officials; the article invents none.