Alibaba Says Qwen 4 Is in Training — 5–10T Models Still a Roadmap, Not a Release

At Apsara Conference in Hangzhou on 22 September 2026, Alibaba put Qwen 4 on the keynote slide — and carefully did not ship it. The official press release says the next-generation model is “currently in training,” with Qwen 4.5 and Qwen 5 projected to scale toward 5 to 10 trillion parameters. Stage chatter named Max/Flash/Plus and a 27B open-weights tier; Alibaba’s published materials do not lock those variant names, so treat them as unconfirmed until a model card appears.

What you can call today is still the Qwen3.8 generation. Alibaba highlighted Recursive Self-Improvement results on Qwen3.8-Max: a month of automated pipeline runs, 33 iterative cycles, and an Artificial Analysis score lift from 40 to 45. A chip-design demo claimed 60-plus hours of self-improvement and a 42% area cut with no performance loss. Impressive theatre — and a reminder that the production API you bill against is not the same object as the roadmap poster.

The multimodal side of the keynote is more immediately product-shaped: Qwen3.8-LiveTranslate (latency down from 2.8s to 2.3s LAAL), Qwen-Audio-3.1 TTS/ASR/realtime updates, Qwen-Image 3.1 later this year for e-commerce creatives, and Qwen Intelligence for agentic phones with HONOR as first partner. That last one matters for SEO adjacent to app ecosystems: when the assistant lives on the lock screen, discovery shifts toward agents that can complete tasks across apps — and toward whatever grounding corpus those agents trust.

For China SEO and marketplace teams, the practical filter is boring and correct. Do not rewrite your content ops around a model with no release date. Do expect Alibaba to keep stuffing Qwen into Taobao, Amap, customer service and phone OEMs. Each of those surfaces can siphon informational queries that used to hit Baidu or classic site search. Watch for API IDs, pricing and citation behaviour when Qwen 4 actually lands — not when the slide says “coming soon.”

Bottom line: Qwen 4 is a training commitment and a competitive signal aimed at DeepSeek, Doubao and the global frontier labs. Until endpoints ship, Qwen3.8-Max plus Model Studio’s web-search tooling remain the stack that can change how your pages get summarised tomorrow morning.

Hardware context from the same keynote underlines that this is a full-stack bet, not a lone LLM drop. T-Head’s Zhenwu V900 accelerator (mass production targeted for Q1 2027), supernode designs aimed at hundreds of thousands of cards, and a 20GW data-centre capacity ambition by 2032 are the industrial backdrop for those 5–10T parameter slides. SEOs do not need to size clusters — but you should read “Qwen 4 in training” as Alibaba committing capex and distribution, which historically precedes deeper embedding of Qwen answers into shopping, maps and customer-service flows that steal informational clicks.

Also note the evaluation theatre around Happyworld Arena and world-model tiers. Brand marketers will see more “AI-generated creative” and simultaneous-interpretation features before they see a clean Qwen 4 API changelog. For search visibility, the nearer-term risk/opportunity is Qwen Intelligence on devices and Model Studio agents with live web search — surfaces that can name or omit your domain long before a 10-trillion-parameter model exists outside a training cluster.

Dan Taylor is the founder and editor of International Search Insights (ISI), the publication behind ReadIsi, and a Partner and Head of Innovation (Search & AI) at SALT.agency. He has worked in search for more than a decade, in-house and agency-side, and is regarded as one of the West’s leading voices on search engines beyond Google, including Yandex, Baidu, Naver and Cốc Cốc. On ReadIsi, Dan covers the engines and AI answer platforms that English-language coverage tends to miss: Baidu and ERNIE, Yandex and Alice AI, Naver, Cốc Cốc, Shenma, Sogou, WeChat and Weibo search, plus DeepSeek and Alibaba’s Qwen. He also keeps dated records of documented changes, such as the Yandex, Baidu and Naver algorithm update histories and the DeepSeek, ERNIE and Qwen model update histories. His approach is to take learnings from global search engines and apply logic with a high evidence threshold. A claim needs a source, patterns seen in one market are treated as a hypothesis until they are checked in another, and where the evidence is thin he says so. For brands, that means advice built on what can be shown, with measurement that separates being cited from being chosen. Dan is credited with coining the term “edge SEO” and won the inaugural TechSEO Boost research competition in 2018 for his work on CDN and edge changes for SEO. In 2026 he was credited with coining “Infinite Tail” to describe how keyword research is changing with AI search. He is the author of AI Search Marketing For The Travel Sector, writes for Search Engine Journal, Search Engine Land and Travel Marketing Guy, and has advised technology, SaaS, travel and enterprise ecommerce brands, including Cloudflare, GitLab and Airbnb. He has given multiple talks and panels at the International Search Summit in Barcelona, and has also spoken at Optimization Moscow, brightonSEO, Google Search Central Live Deep Dive Asia in Bangkok, Athens SEO and the Asia SEO Virtual Summit 2026, where his talk looked at how Baidu, Naver, DeepSeek, Tencent, Cốc Cốc and Yandex built their own AI answers.