Qwen3.8-Flash-Next: Alibaba puts a 125-billion-parameter model on a 6-billion-parameter diet

Alibaba has introduced Qwen3.8-Flash-Next, a multimodal AI model designed to deliver strong coding and workplace performance while using far less computing power than larger rivals. Matthias Bastian reports for The Decoder that the release also previews architectural ideas Alibaba plans to bring to its future Qwen4 model family.

Qwen3.8-Flash-Next contains 125 billion parameters overall, but activates only 6 billion for each token it processes. This is a mixture-of-experts design, which routes each request through only part of the model. Alibaba says this approach helps it outperform its Qwen3.7-Plus predecessor at roughly one ninth of the training cost.

A central change is an N-gram embedding layer with 51 billion parameters. It stores frequently occurring phrases as separate entries, rather than treating every word only in isolation. According to Alibaba, this phrase-level information can remain in standard system memory instead of scarce GPU memory, limiting the added infrastructure cost.

Strong results in coding and office tasks

Alibaba’s published benchmarks show particularly large gains in agentic coding and productivity tests. Flash-Next scores 58.7 on DeepSWE and 62.5 on SWE-bench Pro, both tests of finding and fixing problems in software projects. It also reaches 73.9 on CoWorkBench, compared with 45.1 for DeepSeek-V4-Flash in Alibaba’s comparison.

The model supports a native context window of 262,144 tokens, which can expand to one million tokens through YaRN, a context-extension method. Model weights are available through Hugging Face and ModelScope.

Alibaba plans to offer the production version, Qwen3.8-Flash, through QwenCloud for $0.16 per million input tokens and $0.47 per million output tokens. Those rates are about one twelfth of the listed price for Qwen3.8-Max. The benchmark figures are published by Alibaba and do not necessarily predict performance in every real-world workflow.