Qwen3.8-Flash-Next: 125B open-weights MoE, only 6B active
Alibaba's first Qwen4-architecture preview lands under a community license: 262K context, top-tier coding scores, GGUF and NVFP4 quants already up.
Copy markdown
6B active out of 125B — it fits one box
Qwen3.8-Flash-Next is a 125B-parameter mixture-of-experts that fires just 6B params per token (512 experts, 10 routed + 1 shared). The FP8 checkpoint is ~173 GiB, and community GGUF and NVFP4 builds are already up — so you can self-host it on a single high-VRAM rig instead of a cluster.
Your first look at the Qwen4 stack
It previews Qwen4's architecture: a Gated DeltaNet + Qwen Sparse Attention hybrid at a 3:1 layer ratio, gated residual branches, and a 51B N-gram embedding table for deterministic lookup, trained with the Muon optimizer alongside AdamW. This is where Qwen's next full generation is heading.
Coding and agent scores, not just chat
It posts SWE-bench Pro 62.5, LiveCodeBench v6 91.9, DeepSWE 1.1 58.7 and AndroidWorld 84.5 — ranking #13 of 138 on agentic tasks and top-10 on instruction following. That's a small-active-param model you can point at a real coding agent or a computer-use loop.
262K context, 1M with YaRN
Native context is 262,144 tokens, stretchable to ~1M with YaRN — enough to drop a whole repo or a long agent trace into one prompt. It takes images in and scores MathVision 95.7, so it reads diagrams and screenshots, not just text.
The license catch — read it before you ship
Weights are open under Qwen-Community-1.0, not Apache-2.0, so commercial use needs verification: check the terms before you build a product on it. The official BF16 and FP8 checkpoints are on Hugging Face now; a separate hosted Qwen3.8-Flash on Qwen Cloud is a different model with its own pricing.