Qwen3-Max Released: Alibaba's First Trillion-Parameter Model, Pretrained on 36T Tokens
Alibaba released Qwen3-Max on September 24, 2025 — its largest model ever at over 1 trillion parameters — alongside Nvidia partnership news.
Alibaba Cloud released Qwen3-Max on September 24, 2025: over 1 trillion parameters, its largest and strongest language model to date, pretrained on 36 trillion tokens and entering the global first tier on code, reasoning and agentic tasks.
Per Reuters, Alibaba announced an Nvidia partnership and major data-center plans the same day, lifting its shares. Qwen3-Max runs a dual-track strategy — closed flagship plus open-source matrix — complementing the fully open Qwen3 line.
From Qwen3 (April 2025) to Qwen3-Max (September 2025), Alibaba holds the global lead in open-model downloads while pushing the leaderboard with closed flagships: volume through open source, peaks through closed.
The Business Logic of the Dual Track: Open Source as Funnel, Cloud as Cash Register
Among the giants, Alibaba runs the most complete open-cloud-closed triangle: the fully open, commercially usable Qwen3 matrix built the world's largest derivative-model ecosystem and developer base (see our Qwen3 coverage); when users need more power, Qwen3-Max is available only via the Alibaba Cloud API. Open source is the acquisition funnel, the closed flagship is the monetization exit, and the cloud bill is the actual profit center. It differs from both Meta's ecosystem-play with Llama and OpenAI's pure closed-subscription route — closest to the Android model: give away the operating system, charge for services and the gateway.
The same-day Nvidia partnership and data-center plans completed the other half of the story: a trillion-parameter flagship needs a 10,000-GPU-class cluster behind it, putting Alibaba's compute capex on the same axis as Stargate and the Nvidia-OpenAI deal in the US (see our Stargate coverage) — the other half of frontier competition is always infrastructure.
Our Take
The real story of Qwen3-Max is not the trillion parameters but the business loop it validates — one that runs on cloud, not subscriptions. For anyone choosing a stack, the matrix means a smooth upgrade path: prototype on open Qwen3, switch to the Max API at scale without changing toolchains. For the industry, it proves the trillion-parameter club no longer issues tickets only to US vendors (see our Kimi K2 coverage). Once the open-source volume flywheel spins up, closed rivals face not a single model but an entire ecosystem.
This article aggregates official announcements and public reporting; original sources are linked below.
Source:路透社