Alibaba open-sources Qwen3.8-Flash at about one-ninth training cost

At 11pm Beijing time, Alibaba’s Qwen team put Qwen3.8-Flash online, opening the weights on Hugging Face and ModelScope. It is not just another model in the 3.x line: Alibaba presents it as a technical preview of the next-generation Qwen4 architecture, released early so developers can test the design before the full family arrives.

The model’s central trick is selective computing. Its mixture-of-experts design routes each token to relevant experts instead of running every parameter every time. Qwen3.8-Flash has 125 billion parameters, plus a 51-billion-parameter N-gram embedding module, but activates only about 6 billion parameters per token. Alibaba says training costs about one-ninth those of Qwen3.7-Plus.

That smaller active footprint is paired with a long memory: 262K tokens natively, extendable to 1 million through YaRN. Alibaba’s GDN-plus-QSA attention system compresses historical information and filters important context through a lightweight indexer; on 1-million-token sequences, the QSA kernel reaches reported speedups of up to 7.6x in prefill and 4.9x in decoding.

And so what, concretely? Developers can download the weights, including an FP8 version that lowers the barrier to local deployment, and experiment with long-document processing, agent workloads and batch data handling. For API users, pricing starts at 1 yuan per million input tokens and 3 yuan per million output tokens, which Alibaba presents as roughly a third of V4 Flash; it says output pricing is roughly one-thirtieth that of many frontier overseas models.

The performance claim is also specific, not just cheaper inference: Alibaba reports scores of 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro and 84.5 on AndroidWorld, with stronger coding and office-task results than Qwen3.7-Plus. Those measurements, along with the cost and speed figures, come from the company’s release as reported by Pandaily; no independent test is cited here. The model is therefore best described as a prototype and an invitation to validate Qwen4’s architecture, not as a finished production system.