Alibaba releases Qwen3.8-Max and promises open weights next week

Alibaba has officially released Qwen3.8-Max, describing it as the Qwen family’s most capable model so far. The Qwen team published the model on August 3 and made it available immediately through Qwen Cloud’s API. It also promised to release the model weights next week. That last step would make this the first Max-class Qwen model distributed with open weights, although those files were not yet available at publication time.

Qwen3.8-Max uses a mixture-of-experts architecture with 2.4 trillion parameters in total and 95 billion active for each token. It is based on the Qwen3.5 architecture, accepts text and images, and offers a context window of up to one million tokens. Qwen also documents interfaces compatible with OpenAI and Anthropic APIs, alongside several reasoning-effort settings. Compatibility can make trials easier, but developers should still expect differences in prompts, tool use and output.

The company’s benchmark charts report gains in coding, workplace tasks, research and long-running agent work. Qwen3.8-Max leads several comparisons and trails other frontier models in others. These are vendor-reported measurements, not independent validation, and benchmark scores do not establish reliability in a company’s own workflow.

Alibaba also highlights unusually long autonomous experiments. In one, the model worked for 16 days on a command-line software project. In another, it spent roughly 125 hours reproducing and extending a machine-learning paper. Such demonstrations suggest better persistence across many steps, but controlled projects are not proof that an agent can operate safely without supervision in production.

The release matters because it combines a frontier-scale Chinese model, commercial API access and a stated open-weight plan. Open weights could let cloud providers and large research organisations inspect, adapt and host the model under their own controls. The hardware requirement remains substantial: a 2.4-trillion-parameter model is not a routine local download for most users or small businesses.

For AI users and makers, the immediate option is to test Qwen3.8-Max through the API on real tasks and compare cost, latency and error rates with existing models. The more consequential moment comes when Alibaba publishes the promised files, licence, model card and deployment requirements. Until then, the openness is a commitment rather than a completed release. If the weights arrive as stated, Qwen will increase pressure on frontier-model providers to compete not only on capability and price, but also on how much control customers receive.