Qwen 3.8 Max Weights Released in Open Access

Qwen (Alibaba) has released the weights for the Qwen3.8-2.4T-A95B model — a 2.4 trillion parameter MoE with a hybrid Gated DeltaNet + Gated Attention architecture and up to 1M token context.

Compiled by Sergey KostenchukPublished 2026-08-12Updated 2026-09-18

2026-08-12 Research Google

🤖 Qwen 3.8 Max Weights in Open Access

Qwen (Alibaba) has released Qwen3.8-2.4T-A95B — a Max-class model with open weights. 2.4T MoE parameters (95B active), 92 layers of Gated DeltaNet + Gated Attention, up to 1M context. AIME25: 92-96%, PaperBench: 93.0. vLLM and SGLang support.

🌍 Alibaba's first open weights for a frontier model — available for local deployment and modification.

👤 Run via vLLM or SGLang on B300/MI355X. Quantized versions NVFP4/MXFP4 (~1.2 TB).

Source 1: https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B Source 2: https://vllm.ai/blog/2026-08-12-qwen3.8

Sources