license: other
library_name: gguf
base_model: OpenYourMind/Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated
tags:
- gguf
- qwen
- qwen3
- qwen3.5
- moe
- abliterated
- uncensored
- dpo
- opus
- qwopus
- kimi
- kimi-k2
- multimodal
- vision
- mmproj
- mtp
- q4_k_m
pipeline_tag: image-text-to-text
Support & Community
☕ If these models are useful to you, consider supporting my work — it funds compute for more & larger abliterations.
💬 Discord: discord.gg/rhUZY5GEZr · ₿ Bitcoin: bc1qsvfduzj9fjs9fugpc52yver3f2g8fp7xjxecdv
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated — GGUF
Overview
GGUF build of OpenYourMind/Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated. See the parent repo for the full pipeline: refusal ablation → constrained-LoRA Opus reasoning SFT → unconstrained chosen-completion SFT → Kimi K2.6 reasoning DPO (≈3,000 distilled samples + synthetic data, improving reasoning verbosity on ~12% of requests and removing looping on 2–6% of long-tail conversations).
This repo ships both the language model and the vision projector (mmproj), so it runs as a full multimodal (image + text) model in llama.cpp / LM Studio.
Files
| File | Bits/weight | Size | Notes |
|---|---|---|---|
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated-Q4_K_M.gguf |
~4.6 | ~76 GB | Language model. Q4_K_M keeps output.weight at higher precision. MTP head included. |
mmproj-Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated-F16.gguf |
F16 | ~0.9 GB | Vision projector (qwen3vl_merger, Qwen3.5 vision tower). Load alongside the model for image input. |
Vision (mmproj)
Pass the mmproj file to enable image input. The vision tower is the standard Qwen3.5-122B-A10B Qwen3-VL encoder (carried over unchanged from the base model), F16.
# llama.cpp multimodal CLI
llama-mtmd-cli \
-m Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated-Q4_K_M.gguf \
--mmproj mmproj-Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated-F16.gguf \
--image path/to/image.jpg \
-p "Describe this image." -ngl 99
In LM Studio: keep the mmproj-*.gguf in the same folder as the model — it is detected automatically and the image-attachment button becomes available.
MTP (multi-token prediction)
This build keeps the MTP head (blk.48.nextn.*, qwen35moe.nextn_predict_layers). Recent llama.cpp with qwen35moe MTP support (e.g. LM Studio's llama.cpp 2.15.0) can load it and expose "MTP Speculative Decoding" in the advanced load settings.
⚠️ Caveat: in our testing the MTP head gave no measurable performance gain on this checkpoint. It is shipped for completeness and would need to be retrained to be genuinely useful — happy to do so if there is interest in the model. The model runs fine with MTP speculative decoding off.
Usage (text-only)
Requires a recent llama.cpp build that supports the qwen35moe architecture (Gated DeltaNet linear-attn + MoE).
llama-cli -m Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated-Q4_K_M.gguf \
-p "Explain why the sky is blue." -ngl 99 -c 8192
Hardware
Q4_K_M (~76 GB) + mmproj (~0.9 GB) fits on a single 96 GB GPU, an Apple Silicon machine with ≥ 96 GB unified memory, or CPU + RAM. Leave headroom for KV cache / context.
Notes
- License: Other (inherits from the Qwen3.5 base license)
- Parent (full weights): Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated
- Modality: Text + Vision (via
mmproj) + MTP head (see caveat) - Architecture: Qwen3 MoE (~10B active / 122B total) + Qwen3-VL vision tower
Disclaimer
Use is the responsibility of the user. Ensure your usage complies with applicable laws, platform rules, and deployment requirements.