tags:
- qwen3.8-flash-next
- qwen4-preview
- placeholder
Qwen3.8-Flash-Next-Abliterated
Placeholder — build in preparation. Qwen3.8-Flash-Next (125B-A6B MoE, Qwen4 architecture
preview) is scheduled for open release on 2026-08-26 23:00 UTC+8. This repo will receive
an abliterated (refusal-orthogonalized) bf16 build once three gates clear, in order:
- License review — the upstream license must permit derivative weights.
- Architecture verification — this is a new architecture, not Qwen3_5-dense; the MTP head
layout, linear-attention naming, and the 51B n-gram embedding path all need mapping before any
correct quantization or orthogonalization is possible. Method notes:
nvfp4-mtp-survey. - Toolchain support — llm-compressor / vLLM / llama.cpp support for the new architecture.
No ETA promised. If a correct build turns out to be impossible or the license forbids it, this
page will say so rather than shipping something broken. Track record for what "verified" means
here: Qwen3.8-27B-Abliterated ·
Qwen3.8-27B-Abliterated-NVFP4 ·
Cold-Fusion NVFP4.