If you have read that Qwen 4 already shipped under a different name, you read half the truth. Alibaba confirmed on September 22 that its next model generation is in training, with no official release date. Since then, three dated pieces of engineering evidence and one anonymous model in evaluation have pointed the same direction: the launch is close, and an unannounced Qwen is being tested under a codename.
This article separates confirmed facts from hoaxes. Every claim carries its source and date so you can check them yourself.
1. What's confirmed: four names, no date
At the Apsara Conference in Hangzhou (September 22, 2026), Alibaba announced that Qwen 4 is in training and projected the upcoming family on screen: Qwen4-Max, Qwen4-Flash, Qwen4-Plus, and Qwen4-27B. The slide read "Qwen4 Series Coming Soon" and gave no specs, license, or date.
Two more signals about timing:
- News agency 36Kr reported from the event that Qwen4-27B is the open-weights priority, first in line for downloadable weights.
- A research note from CSC Financial, one of China's largest brokers, dated September 28 concluded that the Qwen 4 release was imminent. It is an analyst assessment, without a specific date.
On October 10, the official developer account on X (@QwenDevs) posted a teaser asking "big or small?". An announcement, not a date.
2. The "Qwen 4 under another name" you can download today
The part of the rumor that is official: on August 26, Qwen released Qwen3.8-Flash-Next, and its own materials describe it as "an early preview of the Qwen 4 architecture". We covered it in our Qwen 3.8 Flash Next analysis: a 125B-parameter MoE with only 6B active parameters per token, 512 experts (10 routed + 1 shared), a 51B n-gram embedding table, and GDN+QSA hybrid attention.
That model is an architecture preview, a rehearsal of the technology the new generation will carry. It is not Qwen 4 and it does not replace Qwen 4: it is the public signal of where the family is heading. Alongside it, the dense Qwen3.8-27B (Apache license) remains the current generation's open-weights workhorse.
3. Kalmia: the ghost model that exposed its family
Here is the most interesting piece of evidence. On October 22, an anonymous model with the codename "Kalmia" appeared on a blind evaluation arena where labs test unannounced models before release.
A researcher ran a tokenizer test with conclusive results:
- Kalmia garbled 6 of 8 tokens that systematically trip up Qwen-family models.
- It garbled 0 of 40 tokens characteristic of other labs.
The test marks the family, not a specific model: other providers also build on Qwen's tokenizer. Still, combined with the engineering evidence below, the reasonable reading is that an unannounced Qwen is in public evaluation. The codename is standard industry practice: OpenAI and Google do the same before every launch.
4. Engineering gives away the calendar
Technical preparations leave public traces, and three have surfaced in recent weeks:
| Date | Trace | What it says |
|---|---|---|
| Oct 23, 2026 | Intel-side pull request to a fork of vLLM Intel GPU kernels | Adds INT4 paths (1-2 GPUs) and FP8 (2 GPUs) for "Qwen4-27B dense". States that no end-to-end run on a real Qwen4 checkpoint was completed |
| Oct 23, 2026 | T-Head (Alibaba's chip unit) issue | Adds a Gated DeltaNet head configuration "needed for Qwen4.0-Max tensor-parallel prefill workloads" (closed Oct 26, not merged) |
| Aug 26, 2026 | Transformers integration | Documents gated residual paths, sparse attention, and n-gram features under the experimental Qwen4-Exp name |
These traces name final model identifiers, which is what distinguishes them from rumors: deployment teams do not prepare kernels for models they do not plan to launch. Several countdown sites estimate an October 30 launch, but there is no official date and that figure is an extrapolation.
5. The hoaxes circulating (and how to verify them)
Every big launch attracts fake benchmarks. This week it's the Qwen family's turn:
| Circulating hoax | What verification says |
|---|---|
| "Qwen 4 Max (Preview) scores 53 on Artificial Analysis" | Absent from the evaluator's leaderboard and posts; the circulating chart has a hand-drawn arrow |
| "Qwen4-Coder reaches 82% on SWE-bench Verified" | The page that published it calls the figure illustrative and cites no primary source |
| "Leaked Qwen 4 repos on ModelScope" | The leaked tables contradict each other on size and ID, with no primary links |
| "Internal testers put it near Opus 5.5" | A subjective forum post reposted with Tieba watermarks |
| "Ships this week" / "lands end of October" | One unsourced X post, repeated by news sites without verification |
The practical rule for any benchmark: if the claim does not link to a public leaderboard, an official repository, or a dated model card, it does not exist. A loose number in a screenshot is not a benchmark.
6. What it means for a business running open-source AI
Beyond the media circus, this launch matters for one reason: Qwen is today the open-weights family with the best capability-per-euro ratio for business automation, and Qwen 4 will move that frontier. Editorial estimates place Qwen4-Max around 55 points on the Artificial Analysis index (versus 45 measured for Qwen3.8-Max), on a leaderboard where the best measured model sits at 58 (Opus 5.5). These are forecasts, not measurements: the real data arrives with the model in hand.
Our operational recommendation:
- Don't migrate anything yet. Your automations on Qwen3.8 keep working, and the launch does not affect them.
- Watch the Qwen4-27B open-weights. It is the leading candidate for cheap local-server models; once its weights publish, measure it against the current 27B and GLM on your actual business use cases, as we did in our open-model comparison.
- Wait for verified benchmarks before spending. The stealth-testing phase (Kalmia) and the prepared kernels indicate a short wait, likely a matter of weeks.
- Ignore leaked benchmarks. None of the circulating ones survives primary-source verification.
One clarification that saves confusion: the 5-10 trillion parameters Alibaba mentions in its roadmap belong to Qwen4.5 and Qwen5, per the Apsara presentation itself. That figure does not describe this generation.
Conclusion: three signals that will tell you Qwen 4 is here
Verification leaves three facts and one expectation. Confirmed: four models announced on September 22, an anonymous model with a Qwen tokenizer fingerprint in evaluation since October 22, and Intel and T-Head kernels naming Qwen4-27B and Qwen4.0-Max since October 23. Expected: a launch within weeks, with the open-weights 27B leading.
The definitive signals will be public and checkable: a Qwen/Qwen4-* repository on Hugging Face, a model ID in Qwen's catalog, or the arena model being unmasked. When they appear, we will measure with real data; until then, leaked benchmarks are noise.
Book your 60-minute technical consultation here and we will assess whether your current automation should wait for Qwen 4 or whether savings are already available with the 3.8 generation on your own server.
Frequently asked questions
When does Qwen 4 come out?
Alibaba has not announced a date. The engineering traces (Intel kernels for Qwen4-27B dated October 23, T-Head's configuration for Qwen4.0-Max) and the evaluation of the anonymous Kalmia model point to an imminent launch, and one estimate targets October 30, but Alibaba has confirmed no date.
What is the "Kalmia" model?
An anonymous model detected on October 22 on a blind evaluation arena. A tokenizer test showed it garbles 6 of 8 tokens characteristic of the Qwen family and 0 of 40 from other labs, indicating it belongs to that family. Alibaba has not confirmed it; the evidence is forensic, not official.
Is Qwen 4 already available under another name?
Partly. What is officially available is Qwen3.8-Flash-Next, which Qwen describes as a preview of the Qwen 4 architecture, plus the current generation's Qwen3.8-27B. The Qwen 4 family (Max, Flash, Plus, 27B) has not launched. The anonymous model in testing suggests the real Qwen 4 is being evaluated, but it is not available.
Is it true that Qwen 4 Max scores 53 on Artificial Analysis?
There is no primary source for that figure. It does not appear on the evaluator's leaderboard or posts, and the circulating chart contains manual edits. Until Artificial Analysis publishes the measurement, that number is a rumor.
Should I wait for Qwen 4 to set up local AI in my business?
It depends on the use case. If you need local capability now, the Qwen3.8-27B (Apache) and Qwen3.8-Flash-Next run today on modest hardware. If your project can wait a few weeks, the open-weights Qwen4-27B may improve the capability-cost ratio, and the moment to decide will be with verified benchmarks and downloadable weights on the table.