The Chinese technology company Alibaba presented a new AI language model, Qwen3.8-Max-Preview, on July 19, 2026, which reportedly includes 2.4 trillion parameters. The company positions it as the second strongest model in the world, directly behind Anthropic’s Claude Fable 5. Independent tests on this are still pending.
Presentation follows two days after Kimi K3
The appearance comes during a phase of intense competition among Chinese AI providers. Qwen announced the new model on the platform X as one of the most powerful systems available today. It is said to be comparable to leading frontier models and is claimed to rank only behind Fable 5. Just two days earlier, the also Chinese company Moonshot AI presented an open model with 2.8 trillion parameters, Kimi K3. Its weights are expected to be released, according to a report by Startup Fortune, on July 27. Qwen3.8-Max-Preview can already be accessed through Alibaba’s cloud subscription Token Plan and the in-house programming tools Qoder and QoderWork. Alibaba also held a stake of around 36 percent in Moonshot AI, according to its fiscal year 2024 annual report cited by the South China Morning Post; the two companies thus remain closely intertwined despite the competition, though a more recent, independently confirmed stake figure is not available given Moonshot’s funding rounds since then. Alibaba has not yet announced a date for the promised release of open weights for Qwen3.8. The move is part of a series of Chinese model announcements in recent weeks, including Zhipu’s GLM-5.2 and other systems from the DeepSeek camp, which together point to a multi-trillion-parameter era among Chinese developers.
Independent tests on performance are still missing
Currently there are no reliable external checks of the claims. Alibaba has not published results on common benchmarks such as SWE-Bench or the LMSYS Arena. The company also gave no figures on how many of the 2.4 trillion parameters are actually activated per request – a metric that determines a model’s real computing costs. In earlier Qwen models, this share was far below the total count: Qwen3-235B-A22B, for instance, activates only 22 of its 235 billion parameters per token. Licensing terms, a model card, and the actual weight files for platforms such as Hugging Face are also still missing. Trade outlets such as Kie.ai point out that real adoption by developers depends on usable licenses, verified weights, and a stable inference stack. A single, self-reported figure is not enough on its own. Observers therefore currently treat the ranking as the world’s second-strongest model as a marketing claim. It can only be confirmed once publicly verifiable tests exist.
Discounts cut costs as with the predecessor
On pricing, Alibaba is following the strategy of its predecessor Qwen3.7-Max, which cost $1.25 per million input tokens and $3.75 per million output tokens – notably cheaper than many Western top models. For Qwen3.8-Max-Preview, Alibaba has not yet announced token-based API prices. Instead, users get access through tiered Token Plan subscriptions. The entry tier costs around six dollars a month, the mid tier about twenty, and the Pro tier around seventy dollars, according to a published price table. Within the Qoder and QoderWork coding tools, Alibaba grants additional discounts on compute quotas: 95 percent during regular hours and up to 99 percent during off-peak windows. By comparison, Moonshot AI reportedly charges $3 per million input tokens and $15 per million output tokens for Kimi K3 through its own API. A direct price comparison between the two models therefore remains only partly possible, since Alibaba has not published a comparable API pricing model.
What matters now is whether Alibaba releases the promised open weights in a timely manner and whether independent labs confirm the claimed ranking behind Fable 5. It also remains unclear how many of the 2.4 trillion parameters are actually active during operation. That will determine how costly running the model is for Alibaba itself, and how affordable access ultimately stays for customers.
