---

license: apache-2.0

tags:

• tensor
• transfer
• gguf
• extract
• mtp
• replace
• graft
• blk
• remove

---

GGUF Tensor Transfer

Stream any tensors you pick from one GGUF file (A — donor) into another

(B — base) to create a new file (C). Selected tensors are copied

byte-exact from A; everything else is streamed from B — no re-quantization,

no full-file RAM load.

👉 Github gguf-tensor-transfer

!image

Features

• Per-tensor selection — the table shows every tensor with its quant

type and size in A, B, and the resulting C. Click ✓ to select, click

column headers to sort, search box filters by name.

• Append donor-only tensors — tensors that exist only in A (e.g. an

extra transformer block like the MTP blk.64 of Qwen3.5/3.6/3.8) are listed

at the bottom and can be added to C; the matching metadata KVs

(block_count, nextn_predict_layers) are synced automatically.

• Shape-safe — a tensor is taken from A only when its dimensions match;

otherwise B's copy is kept and a ⚠ warning is shown.

• Extract → donor GGUF — load just B (Load B (extract only)), select

tensors, and export them to a small standalone donor GGUF.

• Safe write — progress bar with live MB/s, cancel removes the partial

file, and a quick verification runs after the write.