Qwen3.8-Flash-Next: open-weight Qwen4 preview, 6B active, 262K ctx
Alibaba shipped its next-gen architecture as open weights first: hybrid attention, a 51B n-gram table offloaded to RAM, coding scores near the frontier.
Copy markdown
The drop: Qwen4's architecture, shipped open first
Alibaba's Qwen team released Qwen3.8-Flash-Next under the qwen-community-1.0 license — an early preview of the Qwen4 architecture, out as open weights before any flagship. It's a ~125B-total MoE with just 6B active per token across 512 experts (10 routed + 1 shared).
Hybrid attention and a 51B table you park in RAM
It stacks Gated DeltaNet with Qwen Sparse Attention plus gated residuals, and keeps a 51B n-gram embedding table in host memory with async prefetch instead of VRAM. The payoff: 262K native context (YaRN to 1M) while per-token compute stays close to a 6B model.
Coding scores that reach for the frontier
Self-reported: DeepSWE 1.1 58.7, SWE-bench Pro 62.5, SWE-bench Multilingual 81.0, LiveCodeBench v6 91.9, GPQA Diamond 91.7 — up ~4-7 points over DeepSeek-V4-Flash on the SWE tasks. Thinking is on by default; run it on your own harness before trusting the numbers.
Running it: FP8 day-one, but not on a laptop
vLLM, SGLang and TokenSpeed support it out of the gate, with an official FP8 build plus community GGUF and NVFP4 quants. The catch: the full repo is ~360GB in BF16, so even quantized this is workstation-or-server territory, not one consumer GPU — the 6B active count cuts compute, not footprint.
The license catch before you ship
qwen-community-1.0 permits commercial use, but cross 100M monthly users or $20M/month revenue and you must show the model name in your UI; running it as a Model-as-a-Service or 'AI work assistant' business needs a separate license from Qwen. Internal use is exempt.