Qwable-27B-NVFP4-MTP-GGUF

Qwable 27B NVFP4 MTP GGUF

Unofficial NVFP4 MTP GGUF conversion of Mia-AiLab/Qwable-3.6-27b.

The upstream Qwable model is a Qwen3.6 27B fine-tune focused on structured instruction following, coding, and reasoning-style assistant behavior. The upstream model card states that MTP is disabled/not present in the original checkpoint.

This build keeps Qwable as the 64-block base model and grafts a compatible Qwen3.6 27B MTP sidecar as block 64, then converts eligible tensors to NVFP4.

File

Q4_K_M describes the source GGUF lineage. The file itself is an NVFP4 GGUF.

Conversion Summary

Base source:

Qwable-27b_Q4_K_M.gguf

MTP donor:

27B_MTP.gguf

Base source before graft:

qwen35.block_count = 64
qwen35.nextn_predict_layers = <absent>
mtp_tensors = 0

MTP donor:

qwen35.block_count = 65
qwen35.nextn_predict_layers = 1
mtp_tensors = 15

Converted:

433 Q4_K tensors -> NVFP4
65 Q6_K tensors  -> NVFP4
8 Q8_0 tensors   -> NVFP4
360 F32 tensors  -> F32

Validation

Final artifact:

tensor_count = 866
NVFP4 = 506
F32 = 360
qwen35.block_count = 65
qwen35.nextn_predict_layers = 1
mtp_tensors = 15
general.file_type = 39

Confirmed MTP tensor examples:

blk.64.nextn.eh_proj.weight = NVFP4
blk.64.ffn_down.weight      = NVFP4
blk.64.ffn_gate.weight      = NVFP4
blk.64.ffn_up.weight        = NVFP4
blk.64.attn_k.weight        = NVFP4

Attribution

Primary model lineage:

Mia-AiLab/Qwable-3.6-27b

The upstream Qwable repository is marked MIT. Users are responsible for also following the terms of the underlying Qwen/Qwen3.6 lineage and any other applicable upstream components.