Alibaba releases Qwen3.8-Flash-Next, an open-weight preview of a new architecture

Alibaba released open weights for Qwen3.8-Flash-Next, described explicitly as a preview of the architecture underpinning a forthcoming Qwen4 rather than a finished flagship — a different exercise from the 2.4-trillion-parameter Qwen3.8-Max the company had unveiled three weeks earlier.

The model is a 125-billion-parameter mixture-of-experts system activating 6 billion parameters per token, combined with a separate 51-billion-parameter n-gram embedding layer that Alibaba said can run in ordinary system RAM rather than on the GPU, functioning as a large phrase-lookup table the model consults without adding to accelerator memory demands. In an accompanying architecture paper, Alibaba researchers reported that the design — which also combines Gated DeltaNet with global attention layers and uses the Muon optimiser for training stability — matched a 397-billion-parameter predecessor’s downstream results at roughly a ninth of the training cost, using a third of the active parameters and a third of the training tokens.

Alibaba reported benchmark scores including 58.7 on the DeepSWE coding benchmark and 62.5 on SWE-bench Pro, claiming particular strength in coding and office-productivity tasks; these are the company’s own figures. The model natively supports a 262,144-token context window, extendable to one million tokens, and was released under an Apache 2.0 licence on Hugging Face and ModelScope, with a hosted version also priced through Alibaba’s cloud API.

The release continued a pattern in 2026 of Chinese labs pairing large open-weight releases with rapid architectural iteration, disclosing training-efficiency techniques that Western frontier labs had largely stopped publishing.