DavidAU-Qwen3.8-27B-Ultimate-Heretic-Uncensored-8Bit-g32-BF16-MTP-FP16-MTPLX

Overview

This repository provides an optimized MTPLX quantization format of the DavidAU uncensored 27B model, specifically tuned and packaged for high-efficiency inference on unified memory architectures (such as Apple Silicon).

Origin & Credits

• Original Model Creator: DavidAU
• Base Architecture / Model: Qwen 2.7B / 27B variant line
• Quantization & Packaging: Converted to MTPLX format by demetri007

Credit Note: All core model training, fine-tuning, and uncensoring work belong to DavidAU. This release focuses on format conversion and KV/layer optimizations for local runner environments.

Hardware & Performance Context

Tested and verified on:

• Target Hardware: Apple MacBook Pro M1 Max (64 GB Unified Memory)
• Optimization Goal: High tokens-per-second generation rate with minimal KV cache memory overhead, leveraging multi-token prediction (MTP) and hybrid FP16/8-bit layer packing.

Specifications

• Bits: 8-Bit (g32)
• Precision Format: BF16 / FP16 MTP mixture
• Format: MTPLX

Usage Instructions

1. Load into your local inference engine supporting MTPLX files (e.g., LM Studio or llama.cpp builds with MTPLX support).
2. Set context window limits according to available RAM allocations (recommended RAM reservation: ~24–30 GB for standard context sizes).