qwen3.6-27b-mtp-gguf

qwen3.6-27b-mtp-gguf (Enclave model volume)

Curated model volume for Enclave confidential inference: Qwen3.6-27B-MTP at unsloth Dynamic Q4_K_XL (~17.9 GB), a single-file GGUF bundled with the official tokenizer.json so the volume is self-contained.

This is the SELF-DRAFTING variant of qwen3.6-27b-gguf: the GGUF carries the model's trained multi-token-prediction head (one extra layer), and llm-chat's "draft": "mtp" setting uses it for speculative decoding - the head proposes the next few tokens at near-zero cost and the trunk verifies them in one pass. Output is byte-for-byte the trunk's own (exact-match verification); the head only changes speed. No separate draft model, no second weights allocation - prefer this volume over the plain 27b + a draft-model pairing.

Upstream unsloth/Qwen3.6-27B-MTP-GGUF ships many quantizations plus vision mmproj files and no tokenizer.json. This volume carries ONLY the one served quant plus the tokenizer (same rationale as the sibling volumes).

Provenance

The tokenizer is byte-identical to the qwen3.6-27b-gguf / qwen3.5-9b-gguf / qwen3.5-122b-gguf / qwen3.5-0.8b-gguf copies (vocab 248320, <|im_end|> 248046 / <|endoftext|> 248044).

Model geometry (GGUF arch qwen35, hybrid-SSM trunk + 1 dense-attention MTP layer): 64 trunk layers, full_attention_interval 4 (16 trunk KV layers), 4 KV heads, head_dim 256. For VRAM budgeting count the MTP head as one more full-attention KV layer (17 total) - its context is created lazily when a speculative session first opens.

LICENSE is the Apache-2.0 text; both upstreams are Apache-2.0.

Usage on Enclave

Wrapped as a Tinfoil Modelwrap volume (dm-verity; the root hash is part of the enclave measurement). Deployments attach it by name; the wasm-manager mounts it read-only at /models/qwen3.6-27b-mtp-gguf.