At the Yunxi Conference, Qwen4 played four cards at once: Max, Flash, Plus, and a 27B left for you to run locally. This morning, the 2026 Yunqi Conference kicked off in Hangzhou. Alibaba CEO Wu Yongming took the stage to talk about the era of machine intelligence, and then Tongyi Qianwen lead Liu Da Yiheng publicly revealed the Qwen4 family lineup for the first time. Four names: Qwen4-Max, Qwen4-Flash

💡 What You Will Learn

At the Yunxi Conference, Qwen4 played four cards at once: Max, Flash, Plus, and a 27B left for you to run locally. This morning, the 2026 Yunqi Conference kicked off in Hangzhou. Alibaba CEO Wu Yongmi

📜 Table of Contents

• Let's Be Clear: This Is a Teaser, Not a Launch
• On One Hand, Chasing 10 Trillion; On the Other, a 27B for You
• What Qwen4 Changed — You Can Actually Run It Already
• The 27B Line Is the Underrated One
• But Don't Rush to Hype It
• Final Thoughts

Qwen4 Unveils Four Cards at Yunxi Conference: Max, Flash, Plus, and a 27B for You to Run Locally

This morning, the 2026 Yunqi Conference kicked off in Hangzhou. Alibaba CEO Wu Yongming took the stage to talk about the era of machine intelligence, and then Tongyi Qianwen lead Liu Da Yiheng publicly revealed the Qwen4 family lineup for the first time.

Four names: Qwen4-Max, Qwen4-Flash, Qwen4-Plus, Qwen4-27B.

Let's Be Clear: This Is a Teaser, Not a Launch

On-site, there was only one slide about Qwen4 that said "Coming Soon." No parameters, no benchmarks, no pricing, no release date — nothing. The official line is that "Qwen4 based on the next-generation architecture is already in training," and Qwen4.5 and Qwen5 are still in the planning stages.

So if you see claims like "Qwen4 parameters leaked" or "Qwen4 benchmarks revealed" over the next few days, none of it is official.

On One Hand, Chasing 10 Trillion; On the Other, a 27B for You

Liu Da Yiheng put up a parameter curve on stage: the flagship in the Qianwen 2.5 era was 72B, and by Qwen3.8-Max it had reached 2.4T — a 33x increase in two years. Going forward, Qwen4.5 and Qwen5 are set to push to 5 to 10 trillion parameters.

This number has been distorted through several rounds of secondhand reporting. Some English sites wrote it as "5-10T tokens," some translated it as "500 billion to 1 trillion," and some just slapped it onto Qwen4 directly. The original Chinese source says 5 to 10 trillion parameters — roughly 2 to 4 times the current flagship. Alibaba hasn't disclosed how big the model itself actually is.

But on that same lineup there's another extreme: Qwen4-27B, the only open-source model of the four, explicitly aimed at local deployment. The division of labor is clear:

On one side, there are the 10-trillion parameters that only giants can afford to play with; on the other, a 27B left for individual developers. That contrast is the most interesting part of this whole event.

What Qwen4 Changed — You Can Actually Run It Already

Qwen4's architecture isn't a black box. On August 26, Alibaba open-sourced Qwen3.8-Flash-Next, and officially stated it's a technical preview of the Qwen4 architecture.

There are four key changes: attention switched to a hybrid of Gated DeltaNet and sparse attention, the residual stream split into four parallel branches, an N-gram embedding table to move some parameters out of computation, and the optimizer switched to Muon paired with AdamW.

In plain terms: total parameters go up, but the actual compute per token is pinned firmly at 6B. This many-parameters, few-activations approach is exactly the skeleton Qwen4 will inherit.

The 27B Line Is the Underrated One

What got the developer community most excited wasn't Max — it was the 27B.

Two months after release, Qwen3.8-27B became the most-liked model in Hugging Face history, surpassing DeepSeek R1 and Llama. The previous-gen 27B was open-source, supported multimodal, and could run on a single GPU with 17GB of VRAM. Some in the community have already cleared their drives waiting for the quantized version of Qwen4-27B, and are asking whether there'll be a low-resource version for Mac.

Alibaba also showed off some impressive ecosystem numbers this round: over 460 Qwen models open-sourced, over 3 billion global downloads, over 300,000 derivative models, with users including Perplexity, Airbnb, Pinterest, and Reuters.

But Don't Rush to Hype It

Qwen3.8-Flash-Next's weaknesses are equally clear. It beat Claude Opus 4.6 on SWE-bench Pro, but scored only 35.9 on HLE, where Opus 4.6 Max got 40.0; on NL2Repo-Bench it got 48.1, trailing DeepSeek-V4-Flash's 54.2.

In other words, it's genuinely strong on agent and coding tasks, but still falls short on frontier reasoning. Whether that gap can be closed in the official Qwen4 release — there's no answer right now.

Final Thoughts

On the same day, T-Head released its next-gen AI chip Zhenwu V900, with 3x the compute of the previous-gen M890, scalable to 500,000 cards per cluster, mass production in Q1 next year. Wu Yongming's long-term goal is for Alibaba Cloud's global data centers to exceed 20GW by 2032.

Put it all together and Alibaba's thinking is clear: models pushing toward 10 trillion parameters, chips stacking toward 500,000 cards, cloud building toward 20GW. And that Qwen4-27B line is the one left for developers at their own desks.

When the official Qwen4 will arrive, the company hasn't said. But judging by the Qwen3.8 cadence — preview on July 19, release on August 3, open-source weights in mid-August, architecture preview on August 26 — there's a move every two to three weeks on average. This "Coming Soon" probably won't keep people waiting too long.