CoolFace
Modelpublic

OpenYourMind/Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated-MLX-4bit

sourceHugging Faceotherupdated 3mo agoView on Hugging Face
1likes252downloads
Model Card

<div align="center"> <img src="OYM_banner.png" alt="OpenYourMind" width="100%"> </div>

Support & Community

<div align="center">

☕ If these models are useful to you, consider supporting my work — it funds compute for more & larger abliterations.

<a href="https://buymeacoffee.com/oym.kuato" target="_blank"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="48" width="200"></a>

**buymeacoffee.com/oym.kuato**

💬 Discord: discord.gg/rhUZY5GEZr &nbsp;&middot;&nbsp; ₿ Bitcoin: bc1qsvfduzj9fjs9fugpc52yver3f2g8fp7xjxecdv

</div>

Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated — MLX 4-bit

Overview

MLX 4-bit quant of `OpenYourMind/Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated` for Apple Silicon (LM Studio / mlx-vlm). 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).

  • —Multimodal: the vision tower is included and kept in full precision (BF16) — image input works.
  • —Language: 4-bit, group size 64 (MoE routing gates kept at higher precision via the model's quant predicate), ≈ 4.5 bits/weight overall.

MTP

The MTP head is not included in this MLX build: no current MLX runtime (mlx-lm / mlx-vlm, including LM Studio's bundled engine) consumes the qwen3.5 MTP head — they drop mtp.* on load — so it was omitted to keep the model loading cleanly. If/when an MLX runtime adds qwen3.5 MTP support, an MTP-enabled build can be produced from the full weights. (Note: MTP gave no measurable gain in our testing and would need retraining to be useful — see the parent card. For MTP today, use the GGUF build with llama.cpp.)

Files

FileDescriptionSize
model-*-of-00014.safetensors4-bit language weights + BF16 vision tower~65 GB total
config.jsonQwen3_5MoeForConditionalGeneration + quantization (4-bit, g64)—
preprocessor_config.json, video_preprocessor_config.json, processor_config.jsonQwen3-VL image/video processor—
tokenizer*, chat_template.jinja, generation_config.jsonStandard—

Total on disk: ~65 GB.

Usage

bash
pip install mlx-vlm
python -m mlx_vlm.generate \
  --model OpenYourMind/Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated-MLX-4bit \
  --image path/to/image.jpg \
  --prompt "Describe this image." --max-tokens 256

In LM Studio: drop the folder under your models directory (publisher OpenYourMind) and load it with the MLX runtime.

Hardware

~65 GB on disk; needs roughly ≥ 72 GB unified memory to load with usable context. Runs on M-series Macs with 96 GB+.

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 (image / video). MTP not included in this build.
  • —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.