Why this release matters
Large pretrained models can accumulate numerical "training noise" in their weight tensors that harms stability, verbosity and instruction-following. This Hugging Face GGUF release performs post-hoc numerical repair (branded "Genesis") on selected tensors and ships a Hermes agent transfer to produce a locally runnable, uncensored Qwen3.6-35B variant aimed at experimentation with multimodal and agentic workflows without full retraining.
Key Capabilities
- Numeric repair (Genesis): applies custom SVD-based surgery to selected convolutional and FFN tensors to reduce what the author calls "training noise" while preserving learned signal; designed as a post-training, non-gradient intervention.
- Hermes agent transfer: incorporates function-calling style fine-tune blocks derived from the NousResearch/hermes-function-calling-v1 dataset to improve agentic behavior and function-like responses.
- GGUF native build + runtime guidance: packaged in GGUF for llama.cpp / LM Studio / other GGUF runtimes, with recommended APEX quantization and runtime parameter suggestions for common GPUs (e.g., RTX 3060). Vision/mmproj support for image-text inputs is included.
- Model spec & tradeoffs: MoE hybrid 35B with ~3B active parameters per forward, very large native context (claimed 262K), multimodal vision support and multilingual coverage. The model is distributed with an Apache-2.0–noted license on the card.
Who it's for — and tradeoffs
Great fit if you want a locally runnable, uncensored Qwen3.6-derived model to explore multimodal prompts, agentic function-calling, or to test tensor-repair hypotheses without retraining full weights. It’s useful for developers experimenting with GGUF runtimes, quantization profiles, and custom system prompts.
Look elsewhere if you need a production-grade safety-filtered model, strict policy-compliant outputs, or provenance-guaranteed training data. The release purposefully emphasizes "uncensored"/experimental behavior and includes numerical surgery methods that are nonstandard; that increases risk of unexpected behavior, compatibility quirks, and potential safety/legal issues depending on your use.
Practical notes
The author provides recommended quantization strings, runtime flags, and example system prompts; a quantization script reference is shared. The base model and Hermes finetune sources are cited on the model card. This release is primarily oriented to advanced users who can manage local runtime configuration, quantization, and responsible safety controls.
