Focus-Red-MTP

Detached Multi-Token Prediction (MTP) weights for Jaidchen/Focus-Red, derived from Qwen/Qwen3.8-27B.

This repository is a sidecar, not a standalone causal language model. model.safetensors contains only the 15 original mtp.* tensors, unchanged in dtype, shape, and tensor names from the source checkpoint.

• dtype: bf16
• tensors: 15
• parameters: 424 699 392
• tensor bytes: 849 398 784
• MTP layers: 1
• compatible target: Focus-Red / Qwen 3.8 27B text architecture

The corresponding tensors are intentionally absent from the Focus-Red checkpoint.