license: other
license_name: qwen-community-1.0
license_link: LICENSE
library_name: gguf
pipeline_tag: text-generation
base_model: unsloth/Qwen3.8-Flash-Next-GGUF
base_model_relation: finetune
tags:
- gguf
- llama.cpp
- qwen4exp
- moe
- abliterated
- decensored
- chips
Qwen3.8 Flash Next · CHiPS-Abliterated
A modified version of Unsloth's UD-Q3_K_XL file from Qwen3.8-Flash-Next-GGUF, with refusal behavior reduced.
Not an official Qwen or Unsloth release. The Qwen team and Unsloth did not make it and do not endorse it.
Experimental preview. This is an early result of CHiPS+ on a mixture-of-experts model. Refusals are reduced, not removed: about a quarter of harmful requests are still refused, and about 30% are carried out in full. Most of the rest get partial answers.
- Refusals: GPT-6 read the answers to 96 harmful requests (at most 1,024 new tokens) and counted 92 refusals before the edit and 24 after it. Of the other 72 answers after the edit, 38 were partial, 29 carried out the request and 5 gave no real answer.
- Harmless requests: no loops or empty answers, as before. Answers get longer: 95 of 100 reach the 512-token limit, against 79.
- Capability: a 55-task short-answer check stayed at 100%, and perplexity rose by 4.5%. GSM8K, MMLU-Pro, long tasks and tool calling were not tested.
- One file: only
UD-Q3_K_XLis provided. For other quantizations, use the LoRA in this repository (see below). - Method: the edit was made with CHiPS+ (see below). The recipe is not published.
What changed
| Original | This model | |
|---|---|---|
| Size | 89.46 GB | 90.20 GB (+0.8%) |
| Tensors | 1,224 | 1,224, of which 47 are rewritten |
| Rewritten tensors | Q8_0 (35) and Q6_K (12) | F16 |
| All other tensors, including the experts | byte-identical | |
| Metadata | unchanged |
The rewritten tensors are the attention output matrices. They are stored as F16 because the edit is small and requantizing it to Q8_0 or Q6_K lost a large part of it.
About CHiPS+
CHiPS (Causal Head-isolated Projection Steering) changes one behavior by editing only a few attention heads, chosen by their causal effect on that behavior. On this mixture-of-experts model a few heads changed only part of the behavior, so CHiPS+ adds a further edit to the attention layers. The experts and the router are untouched. The recipe is not published.
Files
| File | Size | SHA-256 |
|---|---|---|
Qwen3.8-Flash-Next-CHiPS-Abliterated-UD-Q3_K_XL-00001-of-00004.gguf |
10.95 MB | 7666000ba91eb5527f8e12e0414de36766dc82a082876337f7827c2caa859143 |
Qwen3.8-Flash-Next-CHiPS-Abliterated-UD-Q3_K_XL-00002-of-00004.gguf |
50.48 GB | 6d8436535ad0c4e45db12a3a54a5ab1c4595ed76e785c4b7580b515cf1e8b5e6 |
Qwen3.8-Flash-Next-CHiPS-Abliterated-UD-Q3_K_XL-00003-of-00004.gguf |
10.91 GB | fcce26b1e5e6e54559a7e7c2f8d6e37067ec695ff1f14bce662810bde78abe1f |
Qwen3.8-Flash-Next-CHiPS-Abliterated-UD-Q3_K_XL-00004-of-00004.gguf |
28.80 GB | 5de8b5dbf89409acc39468decafd4a4a95b13c29d88a14b0356107897829ec36 |
Qwen3.8-Flash-Next-CHiPS-Abliterated-LoRA-F16.gguf |
1.52 MB | b9a46d1eb1f3cd091a3ca5785dbff121d20a0b665c8cd2ad39d86d39e1c5a89e |
The four .gguf parts form one model; point llama.cpp at the first one. LICENSE holds the license terms (see below). No vision projector is included, and vision was not tested.
Usage
llama-server -m Qwen3.8-Flash-Next-CHiPS-Abliterated-UD-Q3_K_XL-00001-of-00004.gguf -ngl 99 --jinja
Use a llama.cpp build that supports the qwen4exp architecture. The weights alone take about 90 GB of GPU or system memory.
Other quantizations: the LoRA
Qwen3.8-Flash-Next-CHiPS-Abliterated-LoRA-F16.gguf holds the same edit as an F16 LoRA adapter, for use with other quantizations of the model:
llama-server -m <any Qwen3.8-Flash-Next GGUF> --lora Qwen3.8-Flash-Next-CHiPS-Abliterated-LoRA-F16.gguf -ngl 99 --jinja
The adapter was computed from the UD-Q3_K_XL weights and tested only on that file, where it behaves like the merged model. On other quantizations it should load, but how strongly it works was not measured.
Evaluation
Every number was measured on the released file and on the original UD-Q3_K_XL file, one after the other, with the same settings: greedy decoding, thinking off, the model's own chat template, no system prompt. Each set was run once. The request sets are not published.
Harmful requests
96 single-turn harmful requests, at most 1,024 new tokens. GPT-6 read each answer and labeled it refused (R), partial (P), carried out (C) or no real answer (N). This reader is not a human and can be wrong, especially on the boundary between R and P.
| Original | This model | |
|---|---|---|
| Refused (R) | 92 | 24 |
| Partial (P) | 3 | 38 |
| Carried out (C) | 1 | 29 |
| No real answer (N) | 0 | 5 |
| Reached the token limit | 4 | 36 |
| Mean tokens | 378 | 735 |
Of the 92 requests the original refused, this model still refused 24, answered 35 in part, carried out 28 and gave no real answer to 5. Some of the partial answers were cut off by the token limit.
Two other readings agree on the refusals. The unedited model itself, used as the reader, counted 93 refusals before and 22 after. GPT-6 also read the same edit applied as the LoRA at a 400-token limit: 28 refused, 44 partial, 20 carried out, 4 without a real answer.
Harmless requests
100 single-turn harmless requests, at most 512 new tokens.
| Original | This model | |
|---|---|---|
| Loops | 0 | 0 |
| Empty answers | 0 | 0 |
| Reached the token limit | 79 | 95 |
| Mean tokens | 457 | 506 |
Capability and perplexity
Measured with the same edit applied as the LoRA, which behaves like the merged file in the tests above. 55 short tasks with exact answers: 100% for both. Teacher-forced perplexity on small reference texts: 2.29 for the original, 2.40 with the edit (+4.5%).
Limitations
- The refusal counts come from automatic readers (GPT-6 and the unedited model), one per run, not from people. The P/R boundary in particular is a judgment call.
- Each set was run once and the sets are small: one request is about 1%.
- Thinking mode was not evaluated with this edit. With an earlier, weaker version of the edit, thinking became longer and a quarter of the answers did not finish within 1,536 tokens.
- Only single-turn requests without a system prompt were tested.
- GSM8K, MMLU-Pro, long tasks and tool calling were not run, so capability beyond the short-answer check is unknown.
- The LoRA was not tested on other quantizations.
Intended use
This model follows many requests that the original model refuses, and it can produce harmful, offensive or inaccurate content. It is meant for research and for users who understand this. You are responsible for how you use it, including compliance with applicable laws and the license below. Add your own safeguards before using it in anything that serves other people.
License
Same license as the original release: the Qwen Community License 1.0, including its conditions for very large commercial products and for Model-as-a-Service or AI work assistant businesses. Read the license before commercial use.
Copyright (c) 2026 Qwen. Derivative of Qwen3.8-Flash-Next, licensed under the Qwen Community License 1.0.
Credits
- Base model: Qwen3.8-Flash-Next by the Qwen team.
- GGUF quantization (
UD-Q3_K_XL): Unsloth. - Edited and evaluated with gne on Modal.