Alibaba has unveiled its largest open-source LLM yet and will release Qwen3.8-Max alongside the smaller Qwen3.8-27B next week, expanding the Qwen ecosystem with frontier-scale AI models.
Alibaba has introduced Qwen3.8-Max, the latest addition to its Qwen family of open-source large language models and its most capable model to date. The company confirmed it will open-source the 2.4 trillion-parameter model next week, alongside the smaller and more hardware-efficient Qwen3.8-27B, significantly expanding the Qwen open-source ecosystem.
Qwen3.8-Max is around seven times larger than the Qwen3.5 model released in February. Built using a Mixture-of-Experts architecture, it activates 95 billion parameters while processing a query.
The model supports context windows of up to 1 million tokens, allowing it to analyse more than 200 pages of text or approximately 100 hours of video in a single request. Responses can contain up to 131,000 tokens.
Alibaba has not disclosed the model’s complete architecture. However, previous Qwen releases used the Gated DeltaNet attention mechanism, which offers linear scaling to reduce compute and memory growth when handling long-context workloads. Nvidia has previously said the approach combines the gated update rule and delta rule to improve long-context processing.
In internal evaluations, Alibaba said Qwen3.8-Max completed a 16-day coding project without human input and a chip design optimisation workflow involving more than 500 steps. The model scored 1,668 on the Frontend Code Arena benchmark, placing it 37 points behind Claude Opus 5’s highest configuration while outperforming more than a dozen frontier models, including Meta’s Muse Spark 1.1. Qwen3.8-Max is available immediately through Alibaba Cloud.
