Qwen4-Max-Preview: 4-Day Anonymous Test, Pelican Benchmark, 80% Off — Should You Switch?

Anonymous 4-Day Run Revealed as Qwen4-Max-Preview

On November 7, Qoder introduced an anonymous model codenamed Qoduck without disclosing vendor, parameter count, or benchmarks. Developers quickly suspected it was the unreleased Qwen 4, especially after Qwen's official teaser the day before. On November 11, Qoder confirmed: Qoduck upgraded to Qwen4-Max-Preview , a test build of the new Qwen 4 family. During the anonymous period, the model was upgraded daily ; the version released on November 11 is already a further iteration. A community tester noted it was "much better than when I tested Qoduck on Saturday" — part of that improvement comes from four days of continuous iteration, not a single capability leap. Official guidance: tasks that failed earlier are worth resubmitting.

Why the "Pelican Riding a Bicycle" Test?

The benchmark originated on October 25, 2024, when Django co-founder and Datasette author Simon Willison gave the same prompt to 16 models. It became a famous "folk benchmark" because it stresses four capabilities that standard benchmarks miss:

SVG is code, not drawing. Text models cannot draw but can write code; they must compute coordinates, radii, and Bézier curves for an unseen scene without visual feedback.

Bicycle geometry is hard. Frame, wheels, spokes, chain, cranks, pedals, handlebars — all must occupy correct spatial relationships. Even humans often misdraw a bicycle from memory.

Pelicans cannot ride bicycles. The combination does not exist in reality, so models cannot copy a memorized image; they must truly generate.

Result is instantly judgeable. No Elo score needed; open in a browser to see if it looks right.

By September 2026, the benchmark had accumulated 103 SVGs across 66 models over 23 months , spawning the term pelicanmaxxing — suspicion that vendors secretly fine-tune for this specific prompt. The test used in the article adds two extra constraints: output a single HTML file, and synchronize wheel rotation with pedal motion (cranks, chain, wheels, leg joints moving in phase). This tests physical causal modeling, harder than the original static prompt. A mistranslation in a community screenshot rendered "definitely a large-parameter model" as "绝对是对大多数模型无疑"; the correct reading is the author's subjective judgment, not an official parameter disclosure.

Same Day, Three Conflicting Conclusions — Why They Don't Contradict

The article categorizes three voices by information level:

Community single-run test (Soren @Elieness160708, 11:11 on Qoder CN): "Obvious improvement, definitely a large-parameter model." Caveat: single-turn, single-prompt, single generation; model output is stochastic — each run yields a different pelican.

Other community feedback: "Improvement not huge, picture not good-looking."

Third-party benchmark (GameFeel RE, cited by @follistatindev): Qwen4-Max Preview scores 10.3, below Qwen3.8 Flash/Max. Caveat: this is a game-design sub-task under a specific inference budget; not comparable to coding or frontend tasks. Methodologies differ across leaderboards.

The author argues these test different things. The real test is your own previously failing tasks , not a bird on a bike. Alibaba's own advice: "Put it to the test on a real problem." A high pelican score does not guarantee success on your legacy codebase.

Should You Switch Now?

Availability: Qwen4-Max-Preview is live across Qoder International, Qoder CN, Qoder IDE, JetBrains plugin, Qoder CLI, QoderWake, Cloud Agents, Mobile/Web . Initially personal-only during anonymous phase; now personal, team, and enterprise tiers all have access. Update to latest version and select Qwen4-Max-Preview in the model selector (the old Qoduck entry auto-upgraded).

Pricing: Limited-time 20% of original cost — billing factor reduced from 0.5× to 0.1×, discount auto-applied, end date TBA. Daily 100 universal Credits can be used for this model. (Qoder's X post: "80% off, 0.1x credits".)

Caveats: It remains a preview — no release date for GA, no disclosed parameters or benchmarks, no SLA. For local-first developers: Max is cloud-only, no open weights. The current local option is still Qwen3.8 generation (27B dense, Apache 2.0, ~17GB at Q4). A 27B open-weight Qwen4 is expected based on previous cadence, but not officially promised.

Don't Chase "First Place" — Workflow Lock-In Matters More

Nearly every major lab released a model this month, each claiming a top score. But capability gaps are narrowing, while workflows, project configs, and team collaboration are entrenched in specific tools. A few percentage points on a benchmark rarely justify switching workbenches. Alibaba's strategy: run Qwen 4 preview inside Qoder, gather real-task data anonymously, lower trial barrier with 80% off, then refine for GA. Good for users; vendors compete for the window you open daily.

Author's bottom line: Spend a lunch break this week re-running your previously failed tasks on the preview; but do not restructure production pipelines on a preview model . Wait for GA and open-weight releases before planning migration.

Sources & Verification

Qoder official WeChat post 11-11: "Qoduck Graduates: Upgraded to Qwen4-Max-Preview, Limited-Time 20% Off" (anonymous 4 days, daily upgrades, all products/all tiers, 0.5×→0.1×, daily 100 Credits, professional software focus).

Qoder official X post @qoder_ai_ide: "For 4 anonymous days it was a Qwen test build for Qoder users, updated daily… Limited time: 80% off, 0.1x credits."

Community test tweet: Soren @Elieness160708, 11-11 11:11, Qoder CN, model Qwen4-Max-Preview (self-reported, not official benchmark).

Simon Willison blog 2024-10-25 "Pelicans on a Bicycle", 2025-06 AI Engineer conference talk; Pelican Bicycle Benchmark community data (103 SVG / 66 models / 23 months, as of 2026-09).

Cross-references: 36Kr / PMCAFF 11-11 "Alibaba Raises Ducks, MiniMax Releases Rabbits", Zhixingxing 11-11 Qwen4 announcement, NetEase reprint.

This article uses no unverified parameter counts, benchmark scores, or tok/s data; items marked "projected" or "estimated" are not official facts.