Qwen3.8-Flash-Next introduced as an early preview of Qwen4 architecture

Summary

Simon Willison shares his first-hand impressions of Qwen3.8-Flash-Next, a 125-billion parameter (6 billion active) multimodal MoE model positioned by Qwen as an early architectural preview of Qwen4. He is testing the model on DGX Spark using Unsloth quantized weights.

Why it matters

A 125-billion parameter MoE model with only 6 billion active parameters represents a striking efficiency move that could make frontier-class reasoning capabilities accessible on consumer-grade hardware. Its explicit positioning as a Qwen4 architecture preview offers the community an early window into where Alibaba’s next generation is heading. The release of open weights maintains competitive pressure on closed models.

Highlights

• The 125B total / 6B active MoE design aims for significant performance gains per FLOP
• Runs on DGX Spark with Unsloth quantizations (72.5–79 GB), indicating suitability for local deployment
• Qwen markets this not as a standalone product, but as an early preview of the Qwen4 architecture

Critical take

This is a link share containing clearly preliminary and anecdotal impressions. Since Willison himself notes that he is 'still discovering the model,' there are no comparisons, head-to-head benchmarks, or systematic quality evaluations to support claims of 'significant performance improvements.' Moreover, the results come from heavily quantized checkpoints (IQ1_S, Q2_K_XL), which may mean the observed behavior does not reflect the true capabilities of the full-precision model.