Most large models accumulate low-level numerical noise during training that can degrade stability, verbosity, and long-context behavior. This release applies a post-training numerical ‘Genesis’ repair pipeline (targeted SVD-based denoising and scale corrections) to an uncensored Qwen3.6-A3B 35B MoE base, aiming to restore signal purity without finetuning and to make the model more predictable for long-context and multimodal tasks.
Key Capabilities
- Targeted numerical repair: scans selected tensors (notably ssm_conv1d and FFN blocks) and applies custom SVD + scale corrections to reduce training noise while preserving learned gradients and embeddings; the stated goal is improved stability without retraining. This means fewer saturation/scale-mismatch artifacts in practice.
- MoE multimodal base: retains the original MoE architecture (35B total, ~3B active per forward pass, 256 experts, 40 layers) with native multimodal support (image, text, video) and a very large native context (262K tokens). Useful for experiments that need long context or image-text-to-text pipelines.
- GGUF & runtime compatibility: packaged as GGUF for local runtimes (llama.cpp, LM Studio, koboldcpp); the model author documents recommended quantization/offload settings and APEX quant presets for consumer GPUs.
- Practical additions: includes transferred finetune blocks for the Hermes agent dataset and example assets (SVG/pelican test), plus guidance for chat templates and agent-style system prompts.
Who it's for + Trade-offs
Great fit if you want a locally runnable, uncensored Qwen3.6 variant with improved numerical stability for roleplay, multimodal prompts, or long-context agent experiments and you can manage GGUF runtimes and quantization.
Look elsewhere if you need a safety-filtered or production-compliant hosted model (this is explicitly uncensored and can produce unsafe content), if you require official vendor support, or if you prefer models that underwent conventional finetuning rather than numeric repair. Expect nontrivial compute/VRAM and careful quantization/offload choices for smooth usage.
Where it fits
Use this model as a research/experimental local runtime for multimodal prototypes, agentic function-calling tests, or roleplay experiments where permissive output is acceptable and you need long context and MoE behavior. Prefer larger, curated finetunes or hosted APIs for regulated production deployments.
