Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16

Stamp the Qwen 3.8 sampler: temperature 1.0, top-p 0.95, top-k 20

Model card: library mtplx

Model card: MTPLX 2.11.3 measured speeds, pack quality and search terms

Quantized MTP draft head (INT4/g64 affine, all 8 head matrices)

fix: restore vision tower dropped by the forge convert lane

Model card: FP16 sibling for M1 and M2 Macs

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: model-00003-of-00004.safetensors

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: model-00001-of-00004.safetensors

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: model-00002-of-00004.safetensors

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: model-00004-of-00004.safetensors

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: mtp.safetensors

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: tokenizer.json

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: MTPLX_FP16_CONVERSION_MANIFEST.json

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: model.safetensors.index.json

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: config.json

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: chat_template.jinja

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: mtplx_runtime.json

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: tokenizer_config.json

Qwen3.8-27B MTPLX Optimized-Speed FP16 sibling: generation_config.json

initial commit