This item packages a Qwen3.6-based, MoE 35B model in GGUF format that was numerically repaired ("Genesis‑SVD") to recover signal from corrupted tensors and exposed as a multimodal image-text-to-text and conversational agent. The release targets users who need an uncensored, agentic model variant compatible with common GGUF runtimes and quantization workflows.
What Sets It Apart
- Tensor-level repair (Genesis‑SVD): the author reports reconstructing damaged tensor chunks via SVD-based selection and replacement rather than finetuning, aiming to restore learned signal without retraining. This is positioned as a surgical approach to fix saturated/zeroed blocks and scale mismatches.
- MoE architecture & scale: 35B total params with a Mixture‑of‑Experts routing (256 experts, ~3B active per forward pass), hybrid attention design, very large native context and multimodal capabilities (vision support requires an accompanying mmproj file). The model ships ready-to-run in GGUF and lists runtime compatibility (llama.cpp, LM Studio, koboldcpp).
- Practical readiness: the card includes recommended quantization targets (APEX or Q8_K_P), suggested runtime/decoder settings, and small-resource tuning notes (e.g., RTX 3060 guidance), making it aimed at practitioners who will run local quantized instances.
Who it's for and tradeoffs
Great fit if you: need a locally runnable, uncensored Qwen3.6 variant with multimodal/agent capabilities and want a model already prepared for GGUF-based runtimes and quantization workflows. It appeals to users comfortable evaluating privacy/ethics implications of uncensored models and to those who can manage MoE runtime complexity.
Look elsewhere if you: require a strictly safety-filtered production model, cannot accept legal/ethical liabilities of uncensored outputs, or lack the hardware/software expertise to handle MoE models and large‑model quantization. The “repair” method is a nonstandard intervention on weights—benefits claimed by the author should be validated on your own benchmarks.
Where it fits
Technically positioned as a community-supplied, modified base model: it builds on HauhauCS's uncensored Qwen3.6 base and adapts finetune signals from Hermes datasets. Use this when you want an alternative to official Qwen releases for local experimentation, research into tensor repair techniques, or agent/vision experimentation under an uncensored regime.
