# Alibaba's Qwen3.8-Max Launch Pushes Its Shares Up 6% in Hong Kong > Source: > Published: 2026-08-03 16:12:57+00:00 *Alibaba just released its biggest model yet, a 2.4 trillion parameter system it says trails only Anthropic's Claude. Nobody outside Alibaba has confirmed that.* Alibaba shares jumped as much as 6% in Hong Kong trading on Monday. The company had unveiled Qwen3.8-Max, the full release of a flagship model it only previewed on July 19. According to Bloomberg, the model runs on 2.4 trillion total parameters in a mixture-of-experts architecture, with roughly 95 billion active during any given query. It handles text, images and video. It reads up to a million tokens at once, enough to hold a large codebase or a stack of long documents in a single pass. Alibaba says the model can work almost unsupervised for long stretches. In one internal test the company cited, Qwen3.8-Max spent 16 days building and refining an AI coding tool on its own, writing code, testing it, catching its own errors and fixing them without a human stepping in. That's the kind of claim that sounds impressive until you ask who checked it. So far, the answer is nobody but Alibaba. ## The benchmarks tell a messier story The company's own framing is blunt: second only to Anthropic's Claude Fable 5. On the crowdsourced leaderboard Arena.AI, Qwen3.8-Max is now the top-ranked Chinese model for text generation, though it still sits behind Claude Fable 5 and three Claude Opus variants on the overall board. On multimodal tasks, it ranks second globally, again behind a single Fable 5 variant. That's a real, independently tallied result, and it's the closest thing to outside confirmation the launch has produced. Everything past the Arena leaderboard comes from Alibaba's own numbers. On Terminal-Bench 2.1, Qwen3.8-Max scores 86.6, ahead of both Claude Opus 4.8 and Claude Fable 5 at 84.6, though behind GPT-5.6 Sol's 88.8. But on SWE-bench Pro, it posts 67.7 against Fable 5's 80.0. On FrontierSWE it trails badly, 73.5 to Fable 5's 88.8. Alibaba's launch materials emphasise the wins. They say little about the gaps. Independent trackers haven't weighed in yet. Artificial Analysis hadn't scored the model as of this writing, and neither had the major community leaderboards. The one outside data point that exists is a blind head-to-head from Trilogy AI's StackPerf, where the preview version of Qwen3.8-Max scored 80 against 83 for Moonshot's Kimi K3, another Chinese model released just weeks earlier. If that holds, Qwen3.8-Max isn't just unproven against Claude. It may not even be the best model out of China right now. Developers on Hacker News have been openly sceptical. Their scepticism has a track record behind it. Alibaba has historically shipped detailed benchmark tables alongside launches, as it did for Qwen3.6 and Qwen3.7. This time there's no benchmark table, no model card and no licence attached to the release, and when a lab that usually shows its work suddenly doesn't, that's worth noticing. ## What actually settles this The open-weight promise is the part actually worth watching. Alibaba says it will publish Qwen3.8-Max's weights on Hugging Face and ModelScope next week, along with a smaller Qwen3.8-27B checkpoint. It's billing this as the first time a Qwen-Max-class flagship has gone open source, though every previous Max-tier release from Alibaba, including the original Qwen3-Max, shipped closed. If the weights actually land next week, independent labs can finally run their own evaluations instead of relying on Alibaba's slide deck. Frankly, that's the test that matters here. Not the stock move. A 6% pop in Hong Kong tells you what investors think Alibaba just did. It doesn't tell you whether Qwen3.8-Max can actually do it. This is happening against a backdrop where DeepSeek pushed out V4-Flash and Zhipu is fuelling rumours around GLM-5.5, so Chinese labs are shipping frontier-scale claims almost monthly now. Qwen3.8-Max is the latest and, on paper, the biggest of them. Whether it holds up is a question nobody outside Alibaba's own labs has actually answered yet. **Also read:** [Google's AI Agent Big Sleep Found a Chrome Bug That Predates the iPhone's App Store](https://startupfortune.com/googles-ai-agent-big-sleep-found-a-chrome-bug-that-predates-the-iphones-app-store/) • [Congress presses DoorDash to disclose its use of a Chinese AI model](https://startupfortune.com/congress-presses-doordash-to-disclose-its-use-of-a-chinese-ai-model/) • [MiniMax's H3 Video Model Undercuts Sora and Veo on Price and Openness](https://startupfortune.com/minimaxs-h3-video-model-undercuts-sora-and-veo-on-price-and-openness/)