Alibaba Cloud Open-Sources Qwen3.8-2.4T-A95B, First Max-Level Model Release

Alibaba Cloud's ModelScope community announced that the Qwen team has open-sourced the Qwen3.8-2.4T-A95B model weights, marking the first time a Qwen-Max level model is released. The model has 2.4 trillion total parameters, activates 95 billion per token, and natively supports 262,144 token context, extendable to 1,010,000 tokens.

Qwen3.8 continues the hybrid architecture of Qwen3.5, focusing on improving end-to-end capabilities for coding, office work, research, and long-horizon agent tasks. The cloud version, Qwen3.8-Max, builds on these open weights with additional production features.

Benchmarks show it performs competitively against models like Opus 4.8, Fable 5, and GPT 5.6 Sol, with strong results on PaperBench and IFBench. It uses a MoE architecture with 512 experts per layer, activating 10 routed and 1 shared expert per token, plus multi-step MTP training for predicting multiple tokens.

The post-training strategy includes expanding real-world environments, building a unified reward system, and an online data balancer to support stable RL scaling.