license: other
license_name: swift-open-license-1.0
library_name: gguf
pipeline_tag: text-generation
base_model: d0xin/Swift-Qwen3.8-27B-Uncensored-BF16
tags:

• gguf
• llama.cpp
• qwen3_8
• ternary
• turboquant
• tq2_0
• uncensored
• mtp

ATX-Swift-Qwen3.8-27B-Uncensored-TQ2_0-MTP-GGUF

Ternary 2-bit quantization (TQ2_0) with native Multi-Token Prediction (MTP) draft head for d0xin/Swift-Qwen3.8-27B-Uncensored-BF16.

Overview

• Quant Format: TQ2_0 (TurboQuant 2.06 bpw ternary, ~2.39 BPW container with critical tensors preserved). TurboQuant applies randomized Walsh-Hadamard transform (WHT) rotations dynamically during quantization to suppress outlier channels and retain coherence on dense weights without retraining.
• MTP Speculative Decoding: Native layer 64 (blk.64.*) fused directly into the model (qwen35.block_count: 65). Ready for self-speculative decoding in llama.cpp and llamAmpere without needing an external drafter model.
• Size: 8,167,151,488 bytes (7.61 GiB)
• SHA256: d8165afb60195352f417681fa6348acbfb743c9e6bdc7ba87b9bf1a1e37ea086

Usage with llama.cpp / llamAmpere

llama-server -m ATX-Swift-Qwen3.8-27B-Uncensored-TQ2_0-MTP.gguf   -ngl 99 -fa on -c 32768   --spec-type draft-mtp --spec-draft-n-max 3 --spec-draft-p-min 0.45