rf-detr.cpp Model Manifest

The five detection variants were generated 2026-05-27 from rfdetr 1.7.0 checkpoints. All six segmentation variants were re-generated 2026-07-30 from rfdetr 1.9.0 checkpoints, after the segmentation-head block-count fix. Every model is converted with scripts/convert_rfdetr_to_gguf.py (F32) and re-quantized in-place by the C++ quantizer (build/bin/rfdetr-cli quantize). Re-running the steps below today converts from rfdetr 1.9.0 throughout, since that is what scripts/requirements.txt pins.

The .gguf files themselves are gitignored; this manifest is the canonical record of what should exist and at what size. To rebuild from scratch:

# 1. Convert PyTorch -> F32 GGUF (one per variant)
for v in nano small base medium large \
         seg-nano seg-small seg-medium seg-large seg-xlarge seg-2xlarge; do
    scripts/convert_rfdetr_to_gguf.py --variant "$v" --dtype f32 \
        --output "models/rfdetr-${v}-f32.gguf"
done

# 2. Materialize F16, Q8_0, Q4_K for every F32 source
scripts/build_all_quants.sh

All 44 models were verified to load and run rfdetr-cli detect without error at the time they were generated, and all 44 are current: the 16 stale segmentation files described in earlier revisions of this manifest have been replaced.

Detection variants

Variant	F32	F16	Q8_0	Q4_K
Nano	112.7 MB	60.5 MB	36.0 MB	29.7 MB
Small	119.0 MB	64.0 MB	38.2 MB	31.2 MB
Base	119.2 MB	64.2 MB	38.5 MB	31.5 MB
Medium	125.0 MB	67.2 MB	40.2 MB	32.5 MB
Large	125.9 MB	68.2 MB	41.1 MB	33.4 MB

Segmentation variants

Variant	F32	F16	Q8_0	Q4_K
Seg-Nano	127.1 MB	67.8 MB	39.9 MB	31.8 MB
Seg-Small	127.6 MB	68.3 MB	40.4 MB	32.4 MB
Seg-Medium	133.9 MB	71.6 MB	42.5 MB	33.6 MB
Seg-Large	134.6 MB	72.3 MB	43.1 MB	34.3 MB
Seg-XLarge	141.9 MB	76.7 MB	46.1 MB	36.6 MB
Seg-2XLarge	143.9 MB	78.7 MB	48.2 MB	38.6 MB

Sizes are as measured on the files that are published, in MiB, and they are the post-fix sizes: seg-medium, seg-large, seg-xlarge and seg-2xlarge each carry one or two more DepthwiseConvBlocks than the GGUFs originally published for them (one per decoder layer, so 5 for medium/large and 6 for xlarge/2xlarge, instead of a hardcoded 4). All six segmentation variants have now been re-converted from rfdetr 1.9.0 and re-published, and the loader rejects any surviving copy of an old short-block file with an explanatory error.

A second, unrelated defect lived in the C++ two-stage module rather than in the weights: it omitted upstream's gen_encoder_output_proposals validity mask until commit d256d3e. That mask can only fire on a patch grid of side 50 or more, which means seg-xlarge (grid 52) and seg-2xlarge (grid 64) and no other variant. It changed no GGUF, so nothing here needed re-converting on its account; see docs/conversion.md.

Quant choice notes

• F32: full precision reference, about 120 MB per variant.
• F16: only matmul-multiplicand weights converted; non-matmul tensors (norms, conv kernels, embeddings) stay F32. Loader handles F16 pos_embed via bicubic resample in F32 (see commit 2145c7d).
• Q8_0: best accuracy/size trade for production; about 3x smaller than F32.
• Q4_K: smallest practical quant; rows with ne[0] % 256 != 0 (the decoder's 128-dim MLP halves, 60 tensors) silently fall back to Q8_0 per the C++ quantizer's logic. Net result is still about 3.8 to 4.0x compression.

Heavier K-quants (Q5_K, Q6_K) and the older legacy quants (Q4_0/Q4_1/Q5_0/Q5_1) are supported by rfdetr-cli quantize but are not part of the standard matrix; generate on demand if needed.