Alibaba unveils Qwen3.8-Max, its largest model, ahead of open-weight release

Alibaba unveiled Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model and the largest it had built, making it available first through a hosted API — priced at $2 per million input tokens and $6 per million output tokens — ahead of a promised open-weight release. The model handles text, image and video input and produces text output, with a context window of up to 991,000 input tokens and 131,000 output tokens. Alibaba said it would be the first model in its flagship Max tier to be open-sourced, a departure from its practice of keeping the largest Qwen models closed while releasing smaller variants openly.

Reported benchmark results put Qwen3.8-Max close to, but behind, the leading closed models on general capability: it scored 86.6 on Terminal-Bench 2.1, against 88.8 for GPT-5.6 Sol and 84.6 for Claude Opus 4.8. The larger gains over its own predecessor came on agentic coding tasks — FrontierSWE rose from 40.7 to 73.5 and DeepSWE 1.1 from 21.6 to 56.6 — suggesting Alibaba had focused post-training effort on tool-use and long-horizon coding rather than general knowledge.

The scale of the model complicated the open-weight promise: at 2.4 trillion parameters, the full checkpoint is, in practice, a multi-node datacentre artefact rather than something a typical organisation could run on its own hardware. Alibaba’s smaller Qwen3.8-27B model, released alongside it, was positioned as the realistic option for local deployment. The release continued a pattern through 2026 of Chinese labs matching frontier closed-model benchmarks within months, while pairing the largest releases with open licensing that Western labs had largely abandoned.