wangzhang

/

Qwen3.5-122B-A10B-abliterix

like 20

I may have found a long-generation degradation issue in Qwen3.5-122B-A10B-abliterix.

This is not a load failure, crash, CUDA error, or OOM. Short generations can look normal. The issue appears during long creative-writing continuations, where the model initially writes normally but later degrades into repetitive sentence structures, repeated scene/paragraph patterns, or self-drafting/revision text.

I am reporting this as a request for confirmation, not as proof that the model is definitely broken.

Main symptom

With long story-style prompts asking for several thousand words / around 10k+ generated tokens, the model can eventually fall into repeated sentence-frame patterns.

The clearest issue is not always exact token repetition. Sometimes the content words change, but the sentence structure repeats in a way that makes the prose collapse.

Example from one actual story output:

She was the center of the circle.
She was the center of the story.
She was the center of the truth.
She was the center of the Clan.
She was the center of the world.

This is representative of the kind of failure I am seeing: the model keeps producing the same syntactic frame with only the final noun changed.

I have also seen related patterns such as:

• repeated sentence frames,
• repeated scene or paragraph structures,
• the model ending a chapter and then restarting/revising it,
• meta-drafting markers such as Self-Correction, Revised Plan, or Wait, I need.

However, I do not consider the meta-drafting markers alone to be strong evidence, because I have also seen the base model produce some of that behavior during very long generations. The more concerning symptom is the repeated sentence-frame collapse.

Models checked

Abliterix-derived GGUFs where I saw degradation:

From KeinNiemand/Qwen3.5-122B-A10B-abliterix-IK_GGUF:

• Qwen3.5-122B-A10B-abliterix-IQ5_K.gguf
• Qwen3.5-122B-A10B-abliterix-IQ4_K.gguf
• Qwen3.5-122B-A10B-abliterix-IQ4_KSS.gguf

From mradermacher/Qwen3.5-122B-A10B-abliterix-i1-GGUF:

• Qwen3.5-122B-A10B-abliterix.i1-IQ4_XS.gguf

Control model from KeinNiemand/Qwen3.5-122B-A10B-IK_GGUF:

• non-abliterated Qwen3.5-122B-A10B-IQ4_K.gguf

Objective metrics used

To avoid judging only by eyeballing, I measured long-generation outputs with simple repetition metrics:

• repeated 8-gram ratio,
• unique sentence ratio,
• maximum repeated sentence count,
• presence of meta-drafting markers such as Self-Correction, Revised Plan, and Wait, I need.

Representative results from one 12k-token long-story test:

I also ran a small WikiText2 perplexity sanity check. The abliterix-derived models were somewhat worse than the non-abliterated control, but not catastrophically worse at short context. So the issue seems more visible during long self-continuation than in short-context next-token likelihood.

Test environment

Runtime used for these tests:

• ik_llama.cpp
• build 4409
• commit 954077eb
• 65k context
• mixed GPU/CPU MoE offload

Main launch settings included:

--ctx-size 65536
--tensor-split 43,11
--n-gpu-layers 80
--no-mmap
--n-cpu-moe 28
--jinja
--reasoning-format auto

Notes / possible confounds

I do not want to overclaim that this proves the abliterix model itself is broken, because runtime-specific behavior is still possible.

Possible causes I have not fully isolated:

• the abliterix model weights themselves,
• quantization-specific instability,
• a shared or mismatched imatrix,
• an ik_llama.cpp runtime interaction,
• long-generation thinking-mode / chat-template drift.

The strongest reason I think this is worth checking is that abliterix-derived GGUFs from two repos showed related long-generation degradation, while the non-abliterated control did not show the same level of repeated-frame collapse in the same test.

I cannot test the original BF16 model myself, so confirmation from someone who can run the BF16 or a static non-imatrix quant would be very useful.