Alibaba's Qoder adds free mystery coding model Qoduck, users guess Qwen 4

Alibaba Cloud has added an unnamed “stealth” model called Qoduck to Qoder, its agentic coding platform, and is offering it free every day to individual users for a limited time. The company is not saying what the model is. Its announcement on Nov 7, 2026 read simply “Don’t ask where it came from”, and many early users have concluded that they are testing an unreleased Qwen 4.

A stealth model is one a company ships under a placeholder name, without disclosing which model family or version sits behind it. That approach puts it in front of real developers before any formal launch. Here, it means the most important fact for builders, what Qoduck actually is, remains unconfirmed.

What Alibaba Cloud announced about Qoduck

The announcement is a short post on Alibaba Cloud’s X account, published at 9:00 AM on Nov 7, 2026. In full, it says:

A stealth model just landed in Qoder @qoder_ai_ide: Qoduck. Don’t ask where it came from. Free every day for all individual users, for a limited time.

The post directs readers to download Qoder to try it. It had about 35.5K views at the time the source was captured. It gives no model size, no benchmark results, no architecture details and no end date for the free period.

The account AGTP, which tracks AI releases on X, summarises the situation plainly: “Alibaba has not said what the model is”. AGTP describes its post as a follow-up to a stealth-model story it covered on Nov 4.

How the free Qoduck access in Qoder works

Qoduck is available inside Qoder, not as a standalone API. Per AGTP’s summary, Qoder says new and existing individual users of Qoder International get a daily free allowance, free accounts are eligible, and usage within the daily limit consumes no Qoder credits. Alibaba Cloud’s own post frames this as “for a limited time”.

Users also report that two settings are included at no cost: a context window of up to 1M tokens (the amount of text, measured in tokens, the model can consider at once) and the model’s maximum reasoning level. According to those users, both have to be switched on manually in the model settings. These details come from users rather than an Alibaba specification sheet.

Item	What the sources show	Status
Launch	Nov 7, 2026, inside Qoder	Official
Price to users	Free daily for individual users, limited time	Official
Eligibility	New and existing individual Qoder International users, including free accounts	Per Qoder, as summarised by AGTP
Context window	Up to 1M tokens	User reports
Size	27GB	User claim, unconfirmed
Model identity	Undisclosed	Unconfirmed

Is Qoduck Qwen 4? What users claim and what is confirmed

Replies to Alibaba Cloud’s announcement treat the answer as obvious. One reply reads “we knew it’s qwen 4”, and another thanks the company “for letting us try Qwen 4 for free every day”. A third user writes that Qoduck is “not bad at all for a 27GB model”.

The guesses do not all agree. AGTP collects several competing theories:

• The broad community view is that Qoduck is Qwen 4 or a Qwen 4 variant.
• Some users call it Qwen 4 Max, based on its credit multiplier in Qoder.
• Others say it is Qwen 4 Flash.
• One user, tying it to a 27B size, says it is a 27GB model.
• A few testers wondered whether it is DeepSeek v4.1 Pro instead.

AGTP is explicit: “None of this is confirmed”. It also notes that Alibaba has not released Qwen 4, and that as of Nov 5 there was no model card, download, API name or price for it. For reference, AGTP lists Qwen 3.8 Max as Alibaba’s live model, offered with open weights.

The guesses are also not internally consistent. A model positioned as a “Max” tier and one described as a 27GB model would usually be quite different products, and the sources offer no way to reconcile them. Until Alibaba publishes a model card or names the model, any identification should be treated as speculation.

Early tester reports, and why they are thin evidence

The only performance evidence so far is anecdotal. AGTP gathers these early tester reports:

• One tester says Qoduck’s outputs beat Qwen 3.8 Max.
• Another ran the same prompt through Opus 5.5 and Qoduck and says Qoduck did better.
• Another built a three.js water scene in vanilla JavaScript and reports that Qoduck used far fewer tokens.
• Demand is high enough that responses are slow.

The slowness is echoed in replies to the launch post, where one user calls it “pretty slow due to the crazy demand” and says they have “several benchmarks running in parallel”. No results from those runs appear in the sources.

These reports are single-user impressions on a handful of prompts, not controlled evaluations. No independent leaderboard result and no Alibaba benchmark for Qoduck appears in the sources. A comparison against Opus 5.5 on one prompt says little about general coding quality, and token efficiency on one demo does not predict cost or speed across real workloads. They are reasons to try the model, not reasons to rely on it.

Qoder, and Alibaba’s record of tuning Qwen models for it

Qoder is the vehicle for this launch, so its shape matters. Qoder describes itself as “an agentic coding platform for everyone” and ships as a desktop app, a mobile app, an IDE, a JetBrains plugin and a command-line interface (CLI). Its IDE offers autocomplete, conversational coding and delegated “autonomous programming”. The mobile app lets users monitor agent progress, approve critical actions and review deliverables from a phone. For developers building their own products, Qoder offers an Agent SDK (software development kit) and Cloud Agents APIs, a managed agent platform.

Qoder also supports bringing outside models. Alibaba Cloud’s documentation explains how to add models from Alibaba Cloud Model Studio using an API key, after which they appear in a “Custom” group in the model selector; in the CLI, users confirm Context Window and Thinking Effort settings for each model. That is a separate, paid route from Qoduck’s free allowance, and the documentation does not mention Qoduck.

There is precedent for Alibaba building models specifically for Qoder. In a company blog post, Alibaba Cloud introduced Qwen-Coder-Qoder, “a customized model tailored to elevate the end-to-end agentic coding experience on Qoder,” built on the Qwen-Coder foundation and optimised with large-scale reinforcement learning. In Alibaba’s own evaluation on Qoder Bench, its internal benchmark for real-world software engineering tasks, the company says that model surpasses Cursor Composer-1 in task resolution, with terminal command accuracy on Windows improved by up to 50%. Alibaba also reported production gains: code retention up 3.85%, tool error rates down 61.5% and token consumption down 14.5%.

Those figures are about Qwen-Coder-Qoder, not Qoduck, and nothing in the sources links the two. What the earlier post does show is that Alibaba treats Qoder as a testbed where real user interactions feed back into model training, which it calls a “Model-Agent-Product” flywheel. A stealth release fits that pattern of gathering usage before a public launch, though Alibaba has not said that is the purpose here.

What Qoduck means for builders

For individual developers, the practical offer is simple: a free daily allowance on a model that some testers rate highly, with a reported 1M-token context window. That is enough to compare it against current tools on your own repositories and tasks.

The limits are just as clear:

• No API. Access runs through the Qoder app, not an endpoint. No API name or price exists for Qwen 4 per AGTP’s Nov 5 check, so Qoduck cannot yet be built into a product directly.
• Unknown identity. Without a model card, there is no stated training data, licence, safety evaluation or version stability. A stealth model can change behind the same name without notice.
• Temporary pricing. The free access is limited-time, with no end date given.
• Individuals only. The free allowance is described for individual users of Qoder International. The sources say nothing about team or enterprise plans.
• Slow under load. Testers report slow responses due to demand.
• Data handling. The sources do not say how prompts and code sent to Qoduck are used. Given Alibaba’s stated practice of turning real usage into training signals for Qoder models, teams with sensitive code should check Qoder’s terms before testing on proprietary repositories.

Any evaluation should be treated as a snapshot of an unnamed model on a shared, busy service, not a guide to what a future named release will cost or how it will perform.

What to watch next

The key question is whether Alibaba names the model. A Qwen 4 announcement, with a model card, sizes, licence and API pricing, would turn today’s guesses into specifications. Watch for whether any open-weight release follows, given that AGTP notes Qwen 3.8 Max ships with open weights.

Also watch for structured results. One tester says they are running several benchmarks, and independent leaderboard entries or Alibaba’s own Qoder Bench numbers would give the first comparable evidence. Finally, watch for an end date to the free period and any change in the daily allowance, which will determine how long developers can test Qoduck at no cost.