A model that used to live only behind Alibaba's API is now something you can download and run yourself. For a company that built its reputation on frontier models kept behind a paywall, that is a small earthquake.
In early August 2026 Alibaba's Qwen team released Qwen3.8-Max, and for the first time open-sourced the weights of a "Max-class" model — the most capable tier it had previously reserved for paid API. The open-weight version, Qwen3.8-2.4T-A95B, followed within days. The parameter count is the headline: 2.4 trillion, in a sparse mixture-of-experts design that activates about 95 billion per step, with a 1-million-token context window.
Why "open weight" of a Max model matters
Until now, the pattern was clear: open-weight families (Qwen, DeepSeek, GLM) shipped strong-but-not-flagship models, while the best performance stayed API-only. Qwen3.8-Max breaks that.
- Self-hosting becomes real at the top end. A 2.4T model still needs serious hardware, but institutions can now fine-tune and serve the flagship in-house rather than renting it.
- The ecosystem gets the best teacher. Open weights let vLLM, llama.cpp and university labs build tooling against the strongest Qwen, not a weaker cousin.
- It pressures rivals. If Alibaba gives away the top tier, others must justify keeping theirs closed.
For Chinese developers, the stakes are not only technical. Reliance on US-hosted APIs means exposure to export controls and outage risk; downloadable weights let hospitals, banks and government labs run a frontier-class large model (大模型) entirely on domestic servers. That is why Alibaba's move lands differently in Beijing than in Silicon Valley — it is as much about sovereignty as capability.
The architecture builds on Qwen 3.5, uses hybrid attention, and keeps thinking mode on by default. It is not just bigger; it is the same class of model Alibaba previously sold only through Qwen Cloud. The trajectory is the point: Qwen went from sub-billion edge models to a 2.4T flagship in roughly two years, and each step down the open-weight ladder has pulled the rest of the field with it.
The image side went open too
On 14–20 September 2026 Qwen open-sourced Qwen-Image-2.1, a 7-billion-parameter visual-generation component that unifies text-to-image and image editing in one model, with native transparent (RGBA) output. One caveat: it ships under a research licence, not Apache, so commercial use needs a separate agreement from Alibaba. That nuance matters — "open weight" and "open source" are not the same licence.
China's open-model wave is broad, not just Qwen
Qwen is the freshest move, but the 2026 open-weight push is industry-wide.
- GLM-5 (智谱 / Zhipu): open-sourced 11–12 February 2026 under MIT, scaling to 744 billion parameters (40 billion activated). Zhipu cites SWE-bench-Verified 77.8 and Terminal-Bench 2.0 56.2, and claims the open-source #1 spot on Artificial Analysis. A later GLM-5.2 (June 2026) added a 1-million-token context.
- DeepSeek: the official V4-Pro has been released and, per Huawei and several Chinese reports, achieved Day-0 adaptation to the Ascend platform in 2026 — a sign the "China stack" (model + domestic silicon) is closing loops.
- The direction: the race is shifting from "who has the biggest closed model" to "who gives developers the most usable open weights" — and 2026 is the year that shift became visible across China's model labs, not just Alibaba's.
The unglamorous limits
Open weights do not mean open everything, and bigger does not mean better for you.
- Licence fine print. Qwen-Image-2.1 is research-only; the Max open weights use a custom Qwen licence, not Apache. Check before you ship.
- Hardware reality. A 2.4T MoE is still a datacentre play; the "runs on your laptop" claim belongs to the small variants, not the flagship.
- Benchmarks are vendor-run. SWE-bench and Terminal-Bench scores come from the model makers; independent reproduction lags.
- Evaluation is China-centric. Strong on Chinese; Western third-party testing is thinner, and PRC content rules apply to these models.
What readers can do now
- If you build on open models, test the A95B weights against your own tasks — don't trust a leaderboard number for your workload.
- Read the licence before commercial use, especially for the image model's research-only terms.
- Pair a model with domestic hardware if you are in China — the Day-0 Ascend support for DeepSeek and Qwen signals where the stack is heading.
Honest limitations
Qwen figures come from Alibaba's official Qwen LM blog and the Qwen Cloud changelog. GLM-5 figures come from Zhipu's official research page. DeepSeek V4-Pro is confirmed via DeepSeek's official site but specific launch timing and benchmark details were not independently re-verified here. Benchmark scores (SWE-bench, Terminal-Bench, Artificial Analysis ranking) are vendor-reported. "Open weight" is used precisely: weights are downloadable, but licence terms vary and are not uniformly permissive. This article cites no RMB amounts, so no currency conversion applies. Reporting is current to late September 2026.