license: apache-2.0
tags:
- gguf
- qwen
- qwen4
- flash-next
- mixture-of-experts
- quantized
- mixed-precision
- uncensored
- abliterated
- consumer-gpu
language: - en
- zh
- code
pipeline_tag: text-generation
Qwen4 Flash-Next · XQ, abliterated — reserved, no weights yet
This repository contains no model files and is not ready for inference. It reserves the name for
Qwen4-Flash-Next with the refusal direction edited out of the weights, published in Indra's Mirror's
mixed-precision XQ format, once Alibaba releases it.
Follow the repository to be notified when the files land.
What will be published
| base | Qwen/Qwen4-Flash-Next, the official BF16 release |
| edit | refusal-direction removal applied to the full-precision weights before quantisation |
| format | XQ, mixed-precision GGUF, tiered for a single consumer GPU with system RAM and NVMe behind it |
| reference | an unedited abliterated high-precision GGUF, so every XQ file can be measured against it |
| extras | whatever projector or auxiliary files the official release ships, copied unchanged |
Flash-Next is the mixture-of-experts line, so expect a split, tiered download rather than one file:
the routed experts carry most of the bytes, the dense backbone is a small fraction of it, and some
tables can be read from disk rather than held in memory. The card will say exactly which tier suits
which machine, with the numbers behind each claim.
Status
- Qwen4-Flash-Next released by Alibaba
- weights downloaded and verified against the published hashes
- refusal-direction edit fitted and checked on held-out prompts
- XQ size ladder produced and served
- measurements and model card published
Licence
Follows the Qwen4-Flash-Next release licence. Nothing here overrides it.