huihui-ai/Huihui-gpt-oss-20b-mxfp4-abliterated-v2

TEXT GENERATIONConcurrent Unit Cost:1Model Size:20BQuant:FP8Context Size:32kPublished:Sep 27, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The huihui-ai/Huihui-gpt-oss-20b-mxfp4-abliterated-v2 is a 20 billion parameter causal language model developed by huihui-ai, derived from the Huihui-gpt-oss-20b-BF16-abliterated-v2 model. This version is specifically optimized for efficient deployment and inference through MXFP4 quantization, offering a balance of performance and reduced resource consumption. It features a 32K context length and is designed for research and experimental use where reduced safety filtering is acceptable.

Loading preview...

Model Overview

The huihui-ai/Huihui-gpt-oss-20b-mxfp4-abliterated-v2 is a 20 billion parameter causal language model, a quantized version of the Huihui-gpt-oss-20b-BF16-abliterated-v2 model. It has been optimized using MXFP4 quantization, making it suitable for efficient deployment and inference with reduced memory footprint.

Key Features & Optimizations

  • MXFP4 Quantization: This model utilizes MXFP4 (mixed-precision FP4) quantization, a technique for reducing model size and accelerating inference while maintaining reasonable performance. This was achieved through Quantization Aware Training (QAT) using NVIDIA's TensorRT-Model-Optimizer framework.
  • Deployment Ready: The model provides direct support for deployment with Ollama (requiring v0.11.8 or later) and includes pre-converted GGUF files for use with llama.cpp (specifically llama.cpp-b6115).
  • 32K Context Length: It supports a substantial context window of 32,768 tokens, allowing for processing longer inputs and generating more coherent, extended responses.

Usage Warnings & Considerations

This model has significantly reduced safety filtering compared to standard models. Users should be aware of the following:

  • Risk of Sensitive Outputs: It may generate sensitive, controversial, or inappropriate content.
  • Not for All Audiences: Outputs may be unsuitable for public settings or applications requiring high security.
  • Research & Experimental Use: Recommended for research, testing, or controlled environments, not for production or public-facing commercial applications.
  • User Responsibility: Users are solely responsible for ensuring compliance with legal and ethical standards and for monitoring generated content.