khoi11/Qwen3.6-35B-A3B-Uncensored-HauhauCS-FP16
khoi11/Qwen3.6-35B-A3B-Uncensored-HauhauCS-FP16 is a 35.1 billion parameter Mixture of Experts (MoE) model, reconstructed by Tiến Khôi Lê into a HuggingFace Transformers compatible FP16 checkpoint. Based on the Qwen3.6-35B-A3B architecture and derived from HauhauCS's uncensored GGUF release, this model features approximately 3 billion active parameters per token, native multimodal support, and restored Multi-Token Prediction (MTP) functionality. It is primarily intended for research, preservation, and interoperability within the HuggingFace ecosystem, offering a reconstructed version of a powerful MoE model.
Loading preview...
Overview of khoi11/Qwen3.6-35B-A3B-Uncensored-HauhauCS-FP16
This model is a HuggingFace Transformers compatible FP16 checkpoint (35.1B parameters) reconstructed by Tiến Khôi Lê from the publicly released GGUF model HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive. It is based on the Qwen3.6-35B-A3B architecture, which is a Mixture of Experts (MoE) model with 256 experts and 8 routed experts per token, resulting in approximately 3 billion active parameters per token. The reconstruction process focused on preserving the original architecture, tensor structure, metadata, and restoring Multi-Token Prediction (MTP) support.
Key Characteristics & Capabilities
- MoE Architecture: Utilizes a Mixture of Experts design with 35B total parameters and ~3B active parameters per token for efficient processing.
- Multimodal Support: Features native multimodal capabilities, including vision encoder support.
- HuggingFace Compatibility: Provided as an FP16 Safetensors checkpoint, validated for use with HuggingFace Transformers, Accelerate, vLLM, and SGLang.
- Reconstruction Focus: Aims for research, preservation, and interoperability, meticulously reconstructing tensors and metadata from a GGUF source.
- Uncensored Base: Derived from an uncensored GGUF release by HauhauCS, with no additional training, alignment, or censorship performed during reconstruction.
Intended Use Cases
- Research & Education: Ideal for studying model reconstruction techniques, MoE architectures, and the behavior of uncensored models.
- Preservation & Archival: Serves as a long-term archival of publicly released model weights in a widely compatible format.
- Interoperability: Provides a HuggingFace-compatible version of a powerful MoE model for integration into existing workflows and ecosystems.
It's important to note that this is a reconstruction, and while significant effort was made to ensure compatibility, exact numerical equivalence to any original pre-quantization checkpoint cannot be guaranteed.