File	Quant	Size	Notes
`ettin-reranker-150m-v1-f32.gguf`	F32	572 MB	Full precision
`ettin-reranker-150m-v1-f16.gguf`	F16	288 MB	Half precision (identical scores to f32)
`ettin-reranker-150m-v1-q8_0.gguf`	Q8_0	155 MB	8-bit (scores within 0.06)
`ettin-reranker-150m-v1-q4_k.gguf`	Q4_K	102 MB	4-bit (scores within 0.12)

Quality (vs Python CrossEncoder reference)

Quant	France+	France-	Pasta+	Pasta-
Reference	+11.85	-4.41	+7.79	-5.57
F32	+11.59	-4.11	+7.23	-5.25
F16	+11.59	-4.11	+7.23	-5.25
Q8_0	+11.57	-4.10	+7.19	-5.30
Q4_K	+11.50	-4.13	+7.17	-5.35

Ranking is perfectly preserved across all quantizations.

Per-stage diff harness (f32 GGUF vs PyTorch): all 23 stages PASS, structural gate cos=1.000000, worst per-layer cos=0.999760.

Usage

bash

# Rerank
crispembed -m ettin-reranker-150m-v1-q4_k.gguf --rerank-query "What is AI?" --rerank-doc "Artificial intelligence is..."

# C API
float score = crispembed_rerank(ctx, "query", "document");

# Batch rerank
int n = crispembed_rerank_batch(ctx, "query", docs, n_docs, scores);

Architecture

Base: ModernBERT (22 layers, 768 hidden, 12 heads)
Classifier head: Dense(768→768, GELU) → LayerNorm → Dense(768→1)
Attention: alternating global/sliding (window=128), RoPE theta=160000
Tokenizer: GPT-2 ByteLevel BPE (50368 tokens)

cstr/ettin-reranker-150m-v1-GGUF

Ettin Reranker 150M — GGUF

Files

Quality (vs Python CrossEncoder reference)

Usage

Architecture

cstr/
ettin-reranker-150m-v1-GGUF