On November 7, 2026, Alibaba's coding platform Qoder added a model to its picker called Qoduck. The promotion page in Qoder's own documentation describes it as "a stealth model whose identity has yet to be disclosed," and that is the entire product announcement: no maker, no architecture, no parameter count, no benchmark table, no end date. What the docs do specify is the offer. Individual users on Qoder International get a daily free allowance in which Qoduck's Credit rate drops from 0.5x to 0.0x, even at the highest thinking intensity, and the offer applies automatically, including to free accounts with a zero Credits balance.
Read as marketing, this is a teaser. Read as an experiment design, it is something more interesting: a blind evaluation with the audience as the test set. Alibaba gets weeks of real coding tasks against an unnamed candidate model, users get free compute, and neither side has to attach a reputation to the outcome until the reveal. That framing changes what you should take away from the launch, and it changes how you should test the model yourself.
Context: What Qoder Actually Shipped
Qoder is Alibaba's AI coding tool family (an IDE, a JetBrains plugin, a CLI, cloud agents, and the Qoder International app that this promotion targets). The Qoduck event page is unusually specific about the terms for something with no identity:
| Term | What the docs say |
|---|---|
| Model name | "Qoduck," identity undisclosed |
| Start | November 7, 2026, no announced end date |
| Credit rate inside allowance | 0.0x (normally 0.5x) |
| Thinking intensity | Free "even at the highest thinking intensity" |
| Eligible plans | Free, Pro Trial, Lite, Pro, Pro+, Ultra (Qoder International) |
| Ineligible | Team and Enterprise accounts, which never see Qoduck in the picker |
| Eligible products | Qoder International app only; IDE, JetBrains plugin, and CLI excluded |
| Allowance size | Unpublished; paid tiers get "more free uses" than free accounts |
| After allowance | Continues at the standard 0.5x Credit rate, resets next day |
| Known issue | High demand, slow responses at peak times |
Alibaba Cloud's official X account called it a stealth model and added "Don't ask where it came from." The speculation industry responded instantly: developers have converged on Qoduck being a preview of Qwen 4, and Startup Fortune's coverage lays out the circumstantial case: the 0.5x billing rate, user reports of a 1 million token context window, a Qwen teaser post reading "The becoming begins" on November 6, and an October 28 change merged into Qwen's coding tool repository that sized qwen4 models at a 1 million token context. Chinese AI outlet 53AI goes further, stating flatly that Qoduck "actually corresponds to Qwen 4 Max Preview," and describes the teaser poster as a reference to Zhuangzi's story of the fish whose name is deliberately left blank.
Treat all of that attribution as unconfirmed. Neither Alibaba nor Qoder has confirmed a name, size, or architecture, and the 1 million context figure comes from users, not from any document. What is confirmed is the offer, and the offer is where the information actually lives.
What the Pricing Structure Gives Away
Qoder bills in Credits, an abstraction over tokens that its Credits documentation defines as "the resource quota for AI models," deducted based on the model used and combined input and output token consumption, with failed requests never charged. Every model in the picker carries a factor adjustment against a standard rate. A stealth model sitting at a standard rate of 0.5x, temporarily dialed to 0.0x, tells you two things even before the identity reveal.
First, the serving economics. A 0.5x factor implies Alibaba believes the model costs materially less to serve than the premium models in the same picker. For a first-party model running on Alibaba Cloud infrastructure, marginal inference cost is a scheduling problem, not a per-token invoice from a third party. Qoder has run this play before: the same docs list "DeepSeek-Flash: Free for a Limited Time" and "Qwen3.8-Flash: Free for a Limited Time" as sibling promotions. Free allowances are how Alibaba converts idle capacity and internal-model confidence into distribution.
Second, the target of the subsidy. Free accounts and zero-balance accounts are explicitly eligible, while Team and Enterprise subscribers are excluded entirely. This is not a retention offer for customers; it is an acquisition offer for non-customers. The funnel Qoder is buying is: try the free anonymous model, compare it against whatever you currently pay for, and discover you needed no contract to do it.
Our Read: This Is a Blind Eval, So Run It Like One
Here is the data scientist's reading of the setup. A pre-release model exposed to production traffic, with no name attached and no benchmark published, is a shadow deployment wearing a promotion's clothes. The structure is textbook:
- Anonymity removes reputation risk. A model with no name cannot be blamed for a bad score and cannot be held to a price. Qoder's own docs lean into this: "Try it on your own tasks to see how it performs."
- Real tasks beat toy prompts. Coding agents generate long, messy, repository-grounded workloads: multi-file context, tool calls, refactors with regressions. Early reports collected by 53AI match what a preview checkpoint typically looks like: fast first responses and clean simple completions, with uneven quality on multi-step tasks and repeated edits to existing projects. That is a useful signal about what is being tested, which is exactly the workload class that public benchmarks like SWE-bench only approximate.
- The eval corpus is the prize. Every Qoduck session generates preference data against the models users switch between. Qoder even documents the anti-gaming rule: relaying or reselling access is prohibited and can get your account suspended, which protects the cleanliness of the traffic being collected.
The engineering consequence for you, the person with the free allowance: your usual evaluation habits break here. There is no model card, so no context length to rely on (user-reported 1M is a rumor, not a spec). There is no version pin, so a preview model can change behavior between Tuesday and Friday without warning, and anything you benchmark today may not hold at general availability. There is no output format stability guarantee either; preview-stage routing can change decoding and tool-call behavior silently.
So evaluate cheaply and comparatively. Take the five to ten tasks from your own repository where your current model costs you the most review time: cross-file refactors, test repairs, migrations. Run them against Qoduck inside the daily allowance at maximum thinking intensity, since that is free, and record three things only: does it preserve existing behavior, does it converge or loop, and what did you have to fix by hand. If it's genuinely a Qwen 4 Max preview, this is the cheapest look at a frontier-class coding model available anywhere this month. If it isn't, you spent a daily allowance and learned the same thing about a different model.
One caution from the same playbook we covered in Google's Skimaki stealth coding model: anonymous previews from platform vendors exist to move product decisions, not to give you a production dependency. The reveal, in Google's case as potentially here, may come with different limits, a different rate, or no enterprise availability at all.
Outlook: What to Watch, and What It's Really Testing
Three dates matter. If the Qwen 4 rumor is right, the naming event will likely land near Qwen's own launch cadence, and the reveal page will retroactively tell you what you were actually running. Second, the docs promise advance notice before the promotion ends, so the end-date announcement will reveal how much Alibaba is willing to spend on this data collection; a promotion that runs months is a customer acquisition budget, one that ends at launch was an eval. Third, watch whether Team and Enterprise accounts, deliberately excluded today, ever get access. Their exclusion is the loudest clue that this traffic is for testing, not for selling.
The free token story has a cost chain behind it, and we traced it in From $30K Wafers to Free Tokens: nothing here is free, the payment is your workload data and your switching friction. What makes Qoduck different from past free-token bait, including the kind of discount wars we broke down in Kimi K3.1 Pro's 33 percent cut, is that the subsidy is attached to an unnamed model rather than a named one. Alibaba is not buying share of a comparison table yet. It is buying the comparison table's next row, on your prompts, at zero reputational risk.
If you already use Qoder, the rational move is simple: use the allowance on real work this week, save your diffs and failure cases, and re-run the same set after the identity is revealed. If Qoduck turns out to be Qwen 4 Max Preview and holds its quality at the 0.5x standard rate, the pricing pressure on every other coding model in that picker gets very real, very fast.