From 3 items, 1 important content pieces were selected
Technology News
Technology News
Qwen3.8-Flash-Next: Open-Weight Multimodal MoE Previews Qwen4 ⭐️ 8.0/10
Qwen released Qwen3.8-Flash-Next, an open-weights multimodal mixture-of-experts model that also serves as an early preview of the architecture planned for Qwen4. The model has 125B total parameters but only 6B active per token, which provides a significant inference performance boost. Simon Willison tested it on an NVIDIA DGX Spark using Unsloth's quantized GGUF versions, including the 72.5GB UD-IQ1_S and 78.9GB UD-Q2_K_XL checkpoints, and generated pelican-themed images with both. His favorite result came from the UD-Q2_K_XL quantized model using the xhigh reasoning effort setting.
rss · Simon Willison · Aug 26, 23:52
「Background」 Qwen3.8-Flash-Next is an open-weights multimodal Mixture-of-Experts (MoE) model released by the Qwen team as an early preview of the architecture that Qwen4 will be built on. It has 125 billion total parameters but only 6 billion active per token, which is what gives it a significant efficiency boost. This follows the same pattern as Qwen3-Next, which previously previewed the hybrid Gated DeltaNet and Gated Attention design that later carried through the Qwen3.5, Qwen3.6, Qwen3.7, and Qwen3.8 series.
「Impact」 The release gives self-hosting developers an early, open-weights look at Qwen4's architecture, with Unsloth's quantized GGUF files making the 125B-parameter model practical to run on a DGX Spark.
References
Tags: #Qwen, #open-weights, #MoE, #multimodal, #AI models