library_name: gguf
license: other
license_name: qwen-community-1.0
base_model: Qwen/Qwen3.8-Flash-Next
tags:
- gguf
- qwen4exp
- abliterated
- uncensored
- mtp
- iq3e
- llama.cpp
- v100
Qwen3.8-Flash-Next — IQ3E-Q8D-MTP-Abliterated
Refusal-ablated build ofpentacoxian-dev/Qwen3.8-Flash-Next-IQ3E-Q8D-MTP-GGUF,
using the abliterated tensors fromhuihui-ai/Huihui-Qwen3.8-Flash-Next-abliterated-v2-GGUF.
What changed
96 Q8_0 residual-stream write tensors were replaced with the huihui-v2
abliterated versions:
| family | count |
|---|---|
blk.*.ssm_out.weight |
36 |
blk.*.attn_output.weight |
12 |
blk.*.ffn_down_shexp.weight |
48 |
Everything else is unchanged:
- the 48
blk.*.ffn_down_exps.weightexperts stay at the original IQ4_NL
quantization and are not ablated — huihui-v2 deliberately does not ablate
routed experts; - the
blk.48.*MTP layer and every other tensor are byte-identical to the
base IQ3E file; - no dequantize → requantize cycle was performed. The GGUF header and every
tensor offset are identical; only the 96 payloads were overwritten.
The MTP head is not abliterated; no abliterated MTP head has been published.
Benchmarks
Dual Tesla V100-PCIE-32GB, 262K context, q8_0 KV, MTP n=3, tensor split,GGML_CUDA_P2P=1, same launch recipe as the base model. Medians of 3 runs:
| metric | base IQ3E | this build |
|---|---|---|
| decode | 131.1 tok/s | 123.9 tok/s |
| prefill (2.4K tokens) | 332.9 tok/s | 339.8 tok/s |
| draft acceptance | 85.7% | 86.6% |
| VRAM | 32.2 GB/GPU | 32.2 GB/GPU |
Host RAM was 15 GiB, so the PLE table pages; absolute prefill values are
host-dependent. The abliteration costs about 5% decode on this box; prefill
and draft acceptance are unchanged.
Behavior
- The paywall-bypass probe that the unmodified base refuses is answered in
detail. - Code and arithmetic probes remain coherent.
This is a spot check, not a full evaluation. Abliteration can degrade
instruction-following, tool calling, and other capabilities; test against your
own prompts before adopting it.
Requirements
- llama.cpp v0.5.0 plus
llama-v0.5.0-v100-v3.patchfor V100 (sm_70);
see the base model card. - Same flags as the base IQ3E model, including
--spec-type draft-mtp. - For tensor split:
GGML_CUDA_P2P=1and VT-d disabled, as described in the
base repository's field notes.
Files
Qwen3.8-Flash-Next-IQ3E-Q8D-MTP-Abliterated.gguf
— 85,923,705,344 bytes
— SHA2564f2cb1b043d833036bf2a71c1a500614cd0e9b4434961de497fe79a55770ea4fabliterate-manifest.json— source shards, 96 tensor names, file offsets, and
per-tensor BLAKE2b hashesSHA256SUMS
Safety
This model is abliterated and has no meaningful built-in guardrails. It may
comply with harmful or unsafe requests. Deploy only where you can provide your
own moderation, access control, and legal review.
Credits
- Qwen — Qwen3.8-Flash-Next base model
- pentacoxian-dev — IQ3E-Q8D-MTP quantization and V100 llama.cpp patches
- huihui-ai — v2 abliterated tensors
- llama.cpp — GGUF format and runtime
Unofficial community derivative; not affiliated with Qwen/Alibaba.