MTP / speculative-decoding draft head for gemma-4-12B-it, in the upstream llama.cpp gemma4-assistant GGUF data-plane format (arch gemma4-assistant, with nextn.* (+ masked_embd_* where applicable) tensors and embedding_length_out).

Use it as the drafter for self-speculative decoding against the matching gemma-4-12B-it target:

Quants: F16, Q8_0, Q5_K_M, Q4_K_M, Q4_K_S.

Adapted from google/gemma-4-12B-it-assistant (converted from safetensors). Converted from the official Google safetensors via the AtomicBot converter, then emitted directly in the upstream gemma4-assistant data-plane format. The Gemma license applies.