Same weights, different dtype

How much of a fine-tuned model is already stored if you have its base? I cast the base to fp16, chunked both files with a content-defined chunker at Xet’s chunk sizes, and counted shared chunks.

safetensors as publishedbase BF16, fine-tuned F160.0%

safetensors, both F1627.6%

GGUF F1628.0%

GGUF Q8_028.3%

GGUF Q4_K_M37.8%

Gear chunker at Xet's limits (8 KiB minimum, 64 KiB average, 128 KiB maximum). Base: Qwen2.5-0.5B-Instruct; fine-tuned: the same model after a rank-16 LoRA on every attention and MLP projection, merged and saved as fp16. Source: cdc-chunker README, “Deduplication of model files”.

27.6% matches, and all of it is the token embedding, which LoRA did not train. Every projection matrix changed in about 99% of its values, so no chunk inside one can match.

Two things are worth knowing. First, the dtype decides more than the chunker does. As published, with a BF16 base and an F16 fine-tune, nothing is shared, not even the embedding that never changed, because every byte of a weight differs between the two encodings. Second, quantizing shrinks what can be shared: the bytes saved fall from 278 MB for GGUF F16 to 150 MB for Q8_0 and Q4_K_M.

If the base is already there, store the adapter, 8.8M parameters or about 35 MB in fp32, not the 988 MB merged file.

The per-tensor breakdown, the GGUF numbers and the caveats are in the cdc-chunker README.