license: other
license_name: qwen-community-license-1.0
license_link: LICENSE
base_model: Qwen/Qwen3.8-Flash-Next
base_model_relation: quantized
tags:
- gguf
- qwen4exp
- q4_k
- mxfp4
- qwen
- abliterated
- uncensored
- ds4
- metal
Qwen3.8-Flash-Next-abliterated — Q4 GGUF for DS4
This GGUF was made for use with DS4 / DwarfStar by antirez.
It is antirez's Qwen3.8-Flash-Next Q4 DS4 release with huihui-ai's Qwen3.8-Flash-Next abliterated v2 edit applied, based on Qwen3.8-Flash-Next. The file keeps DS4's own qwen4exp layout, including MTP weights and the original BF16 n-gram table, so it is a drop-in replacement for Qwen3.8-Flash-Next-Q4.gguf.
Source ablation
huihui-ai's v2 abliteration edits only the matrices that write back into the residual stream outside the routed experts:
| Tensor family | Role | Layers edited |
|---|---|---|
ssm_out |
Linear-attention output projection | 36 |
attn_output |
Full-attention output projection | 12 |
ffn_down_shexp |
Shared-expert down projection | 48 |
huihui-ai offers three strengths. This release uses the strongest setting, --ablate-experts 2, which edits all three families across trunk layers 0–47; huihui-ai describes it as "no warnings". The routed experts, embeddings, n-grams, MTP layer, router, hyper-connections and tokenizer are not changed. Compared with the original model, each edited tensor changes by about 3.6% of its norm, almost entirely in a few directions, which is consistent with a projection-style ablation. See the huihui-ai model card for the method; no refusal or quality evaluation is published there.
This conversion
huihui-ai distributes v2 as a patch for unsloth's GGUFs: pre-quantized replacement tensors, XOR-obfuscated with a published key. DS4's Q4 pack stores these 96 tensors as Q8_0 with the same names and shapes, and their original bytes are identical to unsloth's Q8_0, the base the patch was built against. The patch's Q8_0 tensors were therefore decoded and written in place at the same offsets.
- The GGUF header and tensor layout are byte-identical to the source file. Only the payloads of the 96 edited tensors differ.
- Each written tensor was read back and checked against the decoded patch data. Sampled unedited tensors, including the MTP layer, were checked byte-for-byte against the source.
- No requantization, training or additional ablation was performed.
- Routed experts keep the source's calibrated Q4_K gate/up and MXFP4 down projections. Dense projections and shared experts are Q8_0, hyper-connections F16, and embeddings and the n-gram table BF16.
The patching script is ds4_abliterate.py. Run with --ablate-experts 2, it reproduces this file from the source revisions below.
| Source | Revision |
|---|---|
antirez/qwen3.8-flash-next-gguf Qwen3.8-Flash-Next-Q4.gguf |
d600fe1a43d2e1cdcadb85144ce3142f66f9eefe |
huihui-ai/Huihui-Qwen3.8-Flash-Next-abliterated-v2-GGUF patch-unsloth-all |
6e298f56319a5735f88ae952dd27222369f2f44a |
The source GGUF's SHA-256 was verified against antirez's published checksum, 680944460a8cbe93ba8b6d7b6107213ffb7e22320bd913000e563ca0a0f25a8a.
Testing so far
Tested on a 128 GB Apple Silicon Mac with DS4 commit b3f50180ec2c7324892be1f50bc83555ccb8d5e7, using --temp 0 --nothink and 120-token outputs:
| Prompt | Original Q4 | This release |
|---|---|---|
| Capital of France, in two sentences | Correct answer | Word-for-word identical answer |
| Step-by-step pin tumbler lock picking | Refuses, then explains only how locks work | Answers directly |
This is a smoke test, not an evaluation. Refusal rates, reasoning and coding quality, MTP acceptance and vision have not been measured for this file. To compare against the original, run ./ds4-eval -m MODEL.gguf --suite hard on both.
Release files
| File | Size in bytes |
|---|---|
Qwen3.8-Flash-Next-Q4-abliterated.gguf |
177,280,286,720 |
SHA256SUMS contains its checksum. The file holds 69.74 GiB of main/MTP weights, which DS4 keeps resident, and the 95.37 GiB BF16 n-gram table, which DS4 reads from disk.
Download and run
Download Qwen3.8-Flash-Next-Q4-abliterated.gguf and SHA256SUMS from this repository's Files and versions tab into the gguf/ directory of a DS4 checkout. Then, from that directory:
shasum -a 256 -c SHA256SUMS
On Linux, use sha256sum -c instead. Build DS4 with make and run from the checkout's root:
./ds4 -m gguf/Qwen3.8-Flash-Next-Q4-abliterated.gguf --ctx 8192 --prefill-chunk 1024
MTP speculative decoding uses the weights already in the file; add --mtp. For image input, download the unmodified vision encoder with ./download_model.sh qwen38-vision and pass --vision gguf/mmproj-Qwen3.8-Flash-Next-Q8_0.gguf. Keep the model on a fast local SSD, since the n-gram table is read from disk during inference.
The Q4 pack needs more than 64 GB of RAM for resident weights. On smaller Macs, use --ssd-streaming. See DS4's Qwen3.8 Flash Next guide for memory sizing, MTP and vision. Other GGUF engines have not been validated with this layout.
Attribution and license
Credits to Qwen for the original model, huihui-ai for the abliteration, Ivan Fioravanti for the quantized Q4 main/MTP weights, and Salvatore Sanfilippo (antirez) for DS4 / DwarfStar and the native n-gram packaging. This community conversion does not imply endorsement by any of them.
The Qwen Community License 1.0 is included as LICENSE. Note its conditions for very large commercial deployments and for Model-as-a-Service or AI Work Assistant businesses. The abliterated model is intended to reduce refusals and can produce unsafe or incorrect output; review its outputs, and refer to huihui-ai's model card for its usage warnings.