Alibaba's Qwen3.8-Max Arrives With 2.4 Trillion Parameters and Open Weights

What Alibaba announced

On August 3, 2026, Alibaba's Qwen team released Qwen3.8-Max, the largest model it has ever shipped. The headline number is 2.4 trillion parameters, which would make it one of the biggest publicly described language models to date. Alibaba is calling it the new flagship of the Qwen family and has already opened access through an API on its Alibaba Cloud Model Studio platform, with the model also wired into QwenWork, the company's workplace AI product.

The launch matters for a reason beyond raw scale. Alibaba said it will publish the model's weights for public download, the first time it has done that for a Max-class Qwen model. That single decision is what turned a routine flagship update into the AI story of the week.

Inside the model

Qwen3.8-Max is built on the Qwen 3.5 foundation and uses a Sparse Mixture-of-Experts design paired with a hybrid attention mechanism. The 2.4 trillion figure is the total parameter count; only about 95 billion are activated for any given token. That is the trick behind every large MoE model: you get the knowledge capacity of a huge network while paying inference costs closer to a mid-sized one.

The context window runs to 1 million tokens. In practical terms Alibaba says that stretches to a 100-page-plus document, a full television series, or a 100-hour livestream held in memory at once. The model is natively multimodal, handling text, images, and video, and the demos leaned hard on that: web reconstruction from a screenshot, floor-plan-to-3D conversion, animation generation, and short game builds.

None of these capabilities is unique on its own. Frontier context windows and multimodal input have been table stakes since early 2026. What is new is seeing them arrive together, at this size, from a lab that intends to give the weights away.

How it scored

On Alibaba's own benchmark table, Qwen3.8-Max ranks fifth in the Text Arena, second in the Vision Arena, and fourth in the Frontend Code Arena. Those are respectable placements rather than a clean sweep, and the vision result is the standout. The published evaluation scores include 93.0 on PaperBench, 92.1 on OmniDocBench 1.5, 90.4 on VideoMME with subtitles, 86.1 on OSWorld-Verified, and 82.3 on MMMU-Pro. Health and finance reasoning were weaker, at 60.2 on HealthBench and 58.3 on PRBench-Finance.

The more interesting claims are the autonomous ones. Alibaba says the model ran a software-engineering project on its own over 16 days, producing 265 commits and 127 pull requests and building a command-line framework it open-sourced on GitHub. In a separate research task it reportedly worked for 125 continuous hours and wrote more than 7,600 lines of code. It also placed in roughly the 87th percentile of a multimodal dialogue competition, beating 458 of 526 human teams.

Treat all of these as vendor figures until independent evaluations land. Self-reported benchmarks reward the tests a lab chooses to show, and a 16-day autonomous run is far harder to reproduce than a leaderboard score. The open weights, once released, will let outside researchers check the work, which is precisely why that release is the part worth watching.

The open-weight promise

"This also marks the first time we will open-source the weights of a Qwen-Max-class model," the Qwen team wrote. Alibaba said the weights would go up on Hugging Face and ModelScope the week of the launch. As of publication on August 5, 2026, they had not yet appeared, so the most consequential detail of the announcement is still a pending file upload rather than a downloadable model.

If the release lands as described, it changes the competitive math. A downloadable 2.4-trillion-parameter model is not something a hobbyist runs on a laptop, but it is something a well-funded startup, a research lab, or a rival cloud can host and fine-tune without asking Alibaba's permission. That is the lever Qwen has used all year to win developer mindshare against closed American models, and applying it to a top-tier model is a deliberate escalation.

Why the market cared

Investors read the launch as a statement about Alibaba's position in the AI race, not just a product note. Its Hong Kong-listed shares closed roughly 6 to 7 percent higher on August 3, and its U.S.-listed depositary receipts rose in premarket trading, according to market reporting. Alibaba framed the model as competitive with the strongest Western systems, with commentary pitching it against Anthropic's frontier models.

The domestic contest is just as sharp. Qwen3.8-Max is slightly smaller than Moonshot AI's Kimi K3, reported at 2.8 trillion parameters with 104 billion activated, and Alibaba claimed it surpassed Kimi K3 on several tests. Chinese labs are now shipping frontier-scale models on a monthly cadence and competing largely on cost and openness, a strategy that keeps pressure on the pricing and licensing terms of the closed models built in the United States.

What to watch next

Three things will decide whether Qwen3.8-Max lives up to its launch. First, the weights: whether they ship, under what license, and how the full model behaves once anyone can probe it rather than read Alibaba's benchmark table. Second, independent evaluations, especially of the autonomous coding claims that are easy to state and hard to verify. Third, price. Alibaba's edge has come from undercutting Western API rates, and the per-token cost of running a 2.4-trillion-parameter model will tell you how aggressive it is willing to be.

For now, Qwen3.8-Max is the clearest sign yet that the most capable open-weight models may come out of Hangzhou rather than San Francisco. Whether that holds depends on a file upload that, as of this writing, has not happened.