Why this matters
Uncensored model ports like this remove refusal behaviour present in many safety-guarded releases, which reveals the original model's raw generation behavior and lets researchers or red‑teamers probe failure modes and alignment gaps. HauhauCS's release focuses on preserving fidelity through model-specific K_P quantization so you can run near-native-quality inference on consumer hardware while retaining multimodal capabilities.
Key Capabilities
-
Uncensored generation with an "Aggressive" policy — produces outputs without the usual refusal responses. So what? Enables deeper analysis of model behavior, toxic-output testing, and tasks that censored variants may decline, but raises clear safety, ethical, and compliance risks.
-
K_P ("Perfect") quant variants across multiple precisions (Q8_K_P → Q2_K_P). So what? These quant profiles aim to preserve quality where it matters, offering ~1–2 quant levels of effective improvement for a modest size increase, improving local GGUF/llama.cpp inference quality vs. naive quantization.
-
Multimodal support with an accompanying mmproj file. So what? You can perform image→text or image+text tasks locally (LM Studio, llama.cpp, Jan, koboldcpp, etc.) without cloud vision endpoints — useful for offline prototyping of V+L workflows.
-
Large native context and MoE architecture (very large context window advertised). So what? Better for long-form, multi-step reasoning or multi-turn multimodal sessions if your runtime keeps the large context; but it also increases memory and runtime complexity.
Who it's for and trade-offs
Great fit if you are a researcher, red-team engineer, or hobbyist who needs an unlocked local copy of Qwen 3.6‑35B for offline experimentation, fidelity-focused quant testing, or multimodal prototype work. The K_P quant approach is valuable when you want a quality/size tradeoff and compatibility with common GGUF runtimes.
Look elsewhere if you need a safety-hardened model for public-facing products, regulated environments, or any use that requires automated content filtering or compliance guarantees — this release intentionally removes refusal behavior and thus increases downstream moderation burden and legal risk.
Where it sits in the ecosystem
Compared with the official Qwen release, this port prioritizes fidelity and uncensored behavior rather than guardrails. Compared with conservative forks (Balanced variants), the "Aggressive" variant is tuned to maximize output permissiveness. Operationally, expect similar usage patterns to other GGUF-ported foundation models but plan for additional moderation, logging, and containment if you experiment with it.
Practical notes (non-operational)
- Compatibility: loads in llama.cpp, LM Studio, Jan, koboldcpp and other GGUF-compatible runtimes. Vision features require the mmproj file.
- Risk management: apply offline filters, human-in-the-loop review, and strict access controls before letting others use this build.
