Qwen4 Preview: Four Models, One Runs on Your PC

Alibaba has shown its hand on Qwen4 — four model names, a local-PC tier, and a launch window that closes tomorrow — without releasing a single weight, benchmark or price. The preview is real. The launch is not, yet. Here is everything that is actually confirmed, everything that is still rumour, and what to do while you wait.

Quick answer: Alibaba has previewed Qwen4 in four tiers — Max, Plus, Flash and a 27B model built to run on a PC. As of 7 November 2026, no weights, price or date exist.

What Alibaba has actually confirmed

Four names, three dates and one teaser — that is the entire confirmed record on Qwen4. Alibaba has said remarkably little in its own voice, and the little it has said is easy to track:

Date	What happened	Source of the fact
28 August 2026	Qwen3.8-Flash-Next released on GMI Cloud, explicitly described by Alibaba as an experimental preview of the Qwen4 architecture.	Alibaba Cloud release statement
22 September 2026	At the Apsara Conference in Hangzhou, Qwen LLM head Liu Dayiheng announced Qwen4 is in training and arriving “very soon”, naming four tiers: Max, Plus, Flash and 27B.	Apsara keynote
29 October 2026	At Apsara Amsterdam, a slide read “Qwen4 Series Coming Soon” with all four tiers; CTO Feifei Li reportedly told attendees “next week”.	Attendee reports only — not in Alibaba press materials
6 November 2026	The official Qwen account posted a teaser captioned “The becoming begins”.	Official Qwen social account

Notice what is missing from that table: no model card, no licence, no benchmark, no price and no date in writing. As the tracker Newszest put it on 3 November, Alibaba has named four models and published “no weights, no license, no model card, no benchmarks and no price for any of them”. That was true then, and it is still true on 7 November.

The four Qwen4 tiers, one by one

Max is the flagship, Plus the workhorse, Flash the fast one — and 27B is the only tier most people will ever run themselves. The roles below are as Alibaba described them on stage; no specifications have been published for any tier.

Tier	Role as described	What we actually know
Qwen4-Max	Flagship, for the hardest tasks and multi-agent reasoning	Name only. Expected to be API-only, following Alibaba’s past pattern — not confirmed.
Qwen4-Plus	Balanced commercial tier for production pipelines	Name only. Plus has never shipped as open weights in past generations.
Qwen4-Flash	Ultra-low-latency, high-throughput serving	Name only. The August preview, Qwen3.8-Flash-Next, is the early look at this tier’s design.
Qwen4-27B	Local deployment on PCs and private clusters	Described as the open-weights tier — but no weights, licence or model card exist yet.

The 27B tier follows a proven pattern. Qwen 2.5 27B became one of the most-used open models for local deployment, and the Qwen team’s own code commits suggest the family is being prepared the same way: Qwen Code, Alibaba’s terminal coding agent, already references qwen4-max, qwen4-plus, qwen4-coder-plus and qwen4-omni-120b-think identifiers, and a pull request merged on 28 October assigned all Qwen4 identifiers a one-million-token context window — though the commit author noted that figure came from a reviewer and remains unofficial until Alibaba publishes model cards.

Why the 27B model is the one to watch

If Qwen4 matters to you, it is almost certainly because of the 27B tier — the flagship is a cloud API you will rent, while the 27B is a model you could own. And Alibaba has already shown us how that tier is meant to work, through the architecture preview it shipped in August.

Qwen3.8-Flash-Next is a 125-billion-parameter mixture-of-experts model with only 6 billion parameters active per token. The trick is a 51-billion-parameter layer called Parameter-Free Lookup Embedding: instead of doing matrix maths for those parameters, the model looks them up in tables stored in ordinary host RAM (DDR4 or DDR5), so only the 6 billion active parameters need expensive GPU memory. Community testers report the preview running at 21 tokens per second on a single RTX 4090 at a 250,000-token context — a community-reported figure, not an Alibaba benchmark — with a native 262,000-token window stretchable to one million tokens through YaRN interpolation.

That design is the whole point. Traditional dense models need pooled high-end accelerators just to hold their weights in memory; Qwen4’s preview design pushes memorised knowledge into cheap host RAM and keeps only live computation on the GPU. If the final 27B tier keeps that shape, a capable model on a workstation stops being a compromise and starts being the default.

One honest caveat: the preview is not Qwen4, and community tests of it were mixed. In one coding evaluation tracked by the kie.ai deep dive, the Flash-Next preview failed a game-building task five times in a row while the dense Qwen3.8-27B solved it first try. Extreme sparsity buys speed and memory; it may cost consistency on multi-step code. Treat every preview number as a hint about the architecture, not a promise about the finished model.

What is still rumour — and what would prove it

Everything with a date, a price or a benchmark attached to Qwen4 is, as of today, unconfirmed. Here is the rumour list, with the evidence that would settle each item:

A launch on or around 10 November. Community expectation, built on Feifei Li’s reported “next week” comment at Apsara Amsterdam, which would put the window at 2–8 November. Alibaba has published no date. What proves it: weights appearing on Hugging Face, or a post on the Qwen blog.
A one-million-token context window. The Qwen Code pull request of 28 October sets it as the default for qwen4 identifiers — but the commit author flagged the figure as reviewer-supplied and unofficial. What proves it: a published model card.
A “Kalmia” canary model already testing in public arenas. Community trackers report unannounced checkpoints matching the preview’s fingerprints, scoring at frontier level on code tasks. No Alibaba confirmation exists. What proves it: an official acknowledgement, or a model card that matches the observed behaviour.
Recursive self-improvement training and the Muon optimizer. Reported from internal configuration files by community researchers; a 30-day test on Qwen3.8 Max reportedly lifted its Artificial Analysis Intelligence Index score from 40 to 45. What proves it: a technical report from Alibaba.
Open weights for 27B, API-only for Max and Plus. That is Alibaba’s established release pattern, not a Qwen4 commitment. What proves it: a licence file next to the weights on Hugging Face.

The pattern to watch is Alibaba’s own history: the Qwen team has repeatedly posted weights on Hugging Face with no advance notice. A launch may arrive as a quiet repository, not a keynote.

What to do while you wait

The useful move this weekend is not refreshing a countdown — it is running the preview Alibaba already shipped.

Try Qwen3.8-Flash-Next today. Its weights are public, and running it against your own workload is the fastest way to learn how the Qwen4 architecture behaves — its speed, its memory appetite and its failure modes — before the real thing arrives.
Watch the Qwen Hugging Face organisation and the Qwen blog. Those are the two places a real launch will appear first, and Alibaba’s past releases have landed there without warning.
Do not pay anyone for “early access”. No legitimate seller exists before weights are published. Anyone charging for Qwen4 access today is selling a rumour.
Check your RAM, not just your GPU. The preview’s design leans on host memory — 51 billion parameters of lookup tables in DDR4 or DDR5 — so system RAM matters as much as VRAM for local runs.
Set the right expectation. Even if weights land on 10 November, trustworthy benchmarks and independent tests will take days. The launch is the starting gun, not the verdict.

Our take

Qwen4’s preview tells us Alibaba’s priorities more clearly than any benchmark could: cheap local inference first, flagship prestige second. Every other frontier lab is racing to build bigger cloud models; Alibaba spent its preview teaching a 125-billion-parameter model to run on a single gaming GPU by moving memory out of the GPU entirely. That is a bet that the next wave of AI value lives on desks and in private clusters, not only in data centres.

It is also a bet that has not paid off yet. Nothing has shipped, nothing has been measured independently, and the “next week” window closes on 8 November — so this story resolves within days, one way or another. Until then, the honest summary is four names and a working preview. Judge the launch by this checklist, not by the hype:

Is there a Hugging Face repository with downloadable weights?
Is there a licence file stating the terms?
Is there a model card with benchmarks and a context-window figure?
Is there an official Qwen blog post announcing it?

If all four answers are no, nothing has launched — no matter how many countdown posts you see. When the answers turn to yes, we will update this article with the real numbers.

Frequently Asked Questions

Has Qwen4 launched?

No. As of 7 November 2026, Alibaba has previewed and named four Qwen4 models — Max, Plus, Flash and 27B — but has published no weights, licence, model card, benchmark or price for any of them. A launch means downloadable weights or an official announcement, and neither exists yet.

When will Qwen4 be released?

There is no confirmed date. The community expectation of around 10 November 2026 comes from Alibaba Cloud CTO Feifei Li reportedly telling Apsara Amsterdam attendees “next week” on 29 October — an attendee report, not an official announcement. The window that comment implies, 2–8 November, ends tomorrow.

Will Qwen4-27B be free and open source?

It has been described as the open-weights, local-deployment tier, and Alibaba’s past pattern is to release its smaller models openly on Hugging Face. But no licence has been published for Qwen4-27B, so “open” is expected, not confirmed — and the Max and Plus tiers are expected to stay API-only.

Can I run Qwen4 on my own PC?

Not yet — there is nothing to download. The closest thing is Qwen3.8-Flash-Next, the public architecture preview, which community testers report running at 21 tokens per second on a single RTX 4090. The 27B tier is designed for exactly this kind of local use, once its weights are published.

How is Qwen4 different from Qwen3?

Qwen4 introduces a new architecture previewed in Qwen3.8-Flash-Next: a sparse mixture-of-experts design with only 6 billion active parameters per token, plus 51 billion parameters of lookup tables held in host RAM instead of GPU memory. Final specifications, benchmarks and pricing for Qwen4 itself have not been released.

Sources and methodology

Claims in this article were checked on 7 November 2026 against conference reporting and the public trail of the Qwen4 architecture preview. Community-reported figures (arena sightings, local-run speeds, training-method claims) are labelled as such throughout; nothing here rests on hands-on testing, because no Qwen4 weights exist to test. This article will be updated when Alibaba publishes weights or model cards.

kie.ai — Qwen 4 Deep Dive: Architecture and Benchmarks (updated 6 November 2026; the “The becoming begins” teaser, the 125B/6B/51B architecture figures, community RTX 4090 test, Kalmia arena reports)
Emergent — Alibaba Teases Qwen4 Lineup at Apsara, Due Next Week (30 October 2026; the “next week” report, the four tiers, the 28 October Qwen Code pull request)
Newszest — Qwen4: 4 Models Named, 0 Open Weights Confirmed (3 November 2026; the confirmed-vs-unpublished record)
Medium — Qwen 4 Is Stealth Announced (6 November 2026; the official X teaser and the four-tier breakdown)

Internal reading: Kimi K3.1 Pro Open Weights Guide: Use It Without a Server — Kimi’s weights landed on Hugging Face on 4 November, the release pattern Qwen4 watchers are hoping for. Thinking Machines’ Inkling 2 Small — another frontier model that launched closed, with open weights promised only after safety testing. The Best AI in the World Scores Just 12% on Real Work — when Qwen4 benchmarks do arrive, read them by cost per success, not by headline score.

Govind Dheda

Govind Dheda covers AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.