Alibaba Pairs a 10-Trillion-Parameter Qwen Roadmap With a Sharply Cheaper Omni-Modal Model Shipping Now
Alibaba used its annual Apsara Conference in Hangzhou on September 22 to make two announcements that pull in opposite directions. On stage, CEO Eddie Wu and chairman Joe Tsai confirmed that Qwen 4 is in training, with a roadmap toward Qwen 4.5 and Qwen 5 variants scaling to 5 trillion to 10 trillion parameters — up to four times the size of the current 2.4-trillion-parameter Qwen3.8-Max, according to reporting from the South China Morning Post and Reuters. No release date was given for any of the three. Eight days earlier, with far less fanfare, Alibaba's Qwen team had shipped Qwen3.8-Omni-Flash, a multimodal model whose entire pitch is doing more for less.
The chip side of the announcement is where the scale ambitions get concrete. Alibaba's semiconductor unit T-Head unveiled the Zhenwu V900, which it says delivers three times the performance of its predecessor, the M890, and can be linked in clusters of up to 500,000 units. Mass production is scheduled for the first quarter of 2027, positioning the chip as infrastructure for the Qwen 4 generation rather than a product for this year. Alibaba also set a target of more than 20 gigawatts of global data-center capacity by 2032. Investors responded immediately: Alibaba's Hong Kong-listed shares rose 5.1 percent on the day, reaching a one-month high, per Reuters' coverage of the event.
Qwen3.8-Omni-Flash, by contrast, is a shipped, benchmarked product, not a roadmap slide. Built on the open-weighted Qwen3.8-Flash-Next base and released as an API-only model across six cloud regions, it handles text, image, audio and video input with a 1-million-token context window. Alibaba's own evaluations, corroborated by independent write-ups from MarkTechPost and Japan's GIGAZINE, show an average 25 percent improvement across 29 benchmarks versus the prior Qwen3.5-Omni-Plus, with a 36.5-point jump on WildClawBench-MM and an OmniVideoBench score rising from 63.4 to 67.8 while using 45.7 percent fewer tokens. Pricing undercuts the predecessor by roughly 98 percent on audio input and 93 percent on combined audio-visual input, at $0.15 per million input tokens and $0.47 per million output tokens.
Alibaba also disclosed a smaller but telling data point: Qwen3.8-Max's score on the Artificial Analysis intelligence index rose from 40 to 45 through what the company described as self-directed training experiments — models generating and filtering their own training signal rather than relying solely on human-curated data. That claim is unverified by outside benchmarking so far, but it fits a pattern across Chinese labs this year of leaning on synthetic self-improvement loops to close capability gaps without matching the compute budgets of US frontier labs.
"The truly groundbreaking products of the Machine Intelligence era have not yet arrived." — Eddie Wu Yongming, CEO, Alibaba Group
Read together, the two announcements describe a deliberate hedge rather than a contradiction. Alibaba is signaling to investors and to Beijing's chip-independence push that it can credibly compete at frontier scale — hence the 10-trillion-parameter target and the 500,000-chip cluster claim — while its actual shipping cadence stays anchored to cost-per-token improvements that make its models viable for high-volume commercial deployment today. That bifurcation, announcing a scale target with no date attached alongside a real, cheaper release with hard benchmark numbers, is becoming the standard playbook for Chinese labs racing to demonstrate both ambition and commercial discipline in the same news cycle, and it is worth watching whether Qwen 4 actually ships anywhere near the parameter counts teased this week.