Qwen 4 Is 'Very Soon' and the 10-Trillion-Parameter Roadmap Is Out

Alibaba just told the open-weight world something it has been waiting to hear, and it did it on a stage in Hangzhou. At the Apsara Conference in late September, Qwen project lead Liu Dayiheng confirmed that Qwen 4 is already in training on a new-generation architecture and will release "very soon," with the roadmap running out to Qwen 4.5 and Qwen 5 scaling toward 5 to 10 trillion parameters.

For anyone tracking open models, this is the forward-looking signal the recent K2 Horizon and Kimi K3 releases did not give you. Those were snapshots. This is a roadmap, and it says the cost collapse in open-weight models is nowhere near done.

A quick refresh on why this lands hard. Qwen has been one of the most influential open-weight model families of the last two years, the one that runs on consumer GPUs and gets fine-tuned into thousands of downstream tools. When its project lead stands on a stage and says the next generation is already training, that is not a startup teaser. It is one of the largest open model programmes in the world signalling its next move.

What was actually confirmed?

A few things, and it is worth holding a hard line between these and the speculation that surrounds them.

Qwen 4 is in training. That is confirmed, from the project lead, on stage. It runs on a new-generation architecture rather than a point upgrade of Qwen 3, and the release is described as "very soon," which in Alibaba's phrasing has historically meant weeks rather than quarters, though no date was given.

The roadmap is public. Qwen 4.5 and Qwen 5 are projected to scale to 5 to 10 trillion parameters. Note the word projected. That is an ambition and a direction, not a shipped product, and parameter count alone stopped being a reliable predictor of quality a while ago. It tells you where the compute is going, not what the model will do.

The surrounding announcements matter too, because they make clear this is a systems bet, not just a model bet. Alibaba showed custom AI chips, a purpose-built "agentic cloud," and a data centre capacity target of 20 gigawatts by 2032. Qwen3.8-Max, previewed at the same event, reportedly completed 33 automated iterative cycles of its own pipeline work. The picture is a company spending on the whole stack, not just the weights.

One thing the announcement did not fully settle is availability. Alibaba has a long record of open-sourcing much of Qwen, but the frontier flagship models have increasingly shipped via API first. For builders, "very soon" is only useful if it means weights you can actually run, and that part was left unsaid.

What is still rumour?

The 74 percent figure needs a health warning. A public prediction market is pricing a Qwen 4 launch before 1 November at around 74 percent. That is a market price, not a fact. Prediction markets are a useful read on what informed people believe, but they have been wrong about release dates before, and a "launch" can mean a paper, a few API endpoints, or a full open-weight drop. Do not book a sprint around it.

Similarly, "5 to 10 trillion parameters" is easy to read as the headline, but it is the least concrete thing on the slide. No architecture details, no training cost disclosure, no benchmark claims. Treat it as a statement of intent, and treat anything more specific than that, especially anything with a date attached, as unconfirmed until it ships.

Why should builders care now?

Here is the part that matters if you build on open models. You do not need to wait for Qwen 4 to act, and you should not.

The lesson of the last year of open-weight releases is that the leader changes every few months, and the cost of any given capability keeps falling. Qwen 3, Kimi K3, K2 Horizon, and now Qwen 4 in the wings: each one resets the price-to-performance curve, and each one makes whatever you hard-coded last quarter look expensive in retrospect.

The builders who got burned are the ones who wired their product to one model, one provider, one API surface. The builders who did well are the ones who treated the model as a swappable component behind an abstraction layer.

The other reason to care now is cost. Every time a capable open model ships, the price of the equivalent closed capability drops, and your negotiating position with your current provider improves. Builders who can say "I can move to Qwen 4 next week" pay less than builders who cannot. Model-agnosticism is not just an architecture nicety. It is a purchasing strategy.

What does model-agnostic actually mean in practice?

It is less abstract than it sounds, and it is mostly a discipline question, not an engineering one.

First, keep your prompts and your orchestration separate from your model choice. Do not tune prompts to the quirks of one model; tune them to the task, and let a routing layer pick the model. When Qwen 4 drops, you want to swap it in behind a flag, not rewrite your integration.

Second, standardise on an interface. Most serious open-weight and API models now speak the same request shapes: OpenAI-compatible endpoints, tool calling, structured output. Build against that common surface and you keep optionality for free.

Third, keep a benchmark harness warm, not a vendor relationship. The moment a new model ships, the only question that matters is whether it beats what you are running, at your task, for your cost. If you have a small eval set you can run in an afternoon, you answer that in hours. If you do not, you will be reading someone else's leaderboard and guessing.

Fourth, keep the door open to self-hosting. Open weights are the whole point of caring about Qwen rather than a closed API. If your stack assumes a hosted endpoint, you are paying a premium and surrendering the swap. Architect so the same task can run on a hosted model today and a downloaded weight tomorrow, and you are ready for whichever direction the market takes.

The takeaway

Qwen 4 is coming, and the 10-trillion-parameter roadmap behind it is Alibaba telling you, as clearly as a company ever does, that the open-weight cost curve has further to fall. What you do with that is your call, and it should not be to wait. Abstract your model layer, keep a benchmark harness ready, and treat the next release as a pricing event you are prepared to exploit, not a surprise you have to react to. The builders who are ready for Qwen 4 will be the ones who do not need Qwen 4 specifically.