tencent/R3-rerank-0.6b

Primitive: /score · Score · Qwen3

The latest agent skill reranking model at the 0.6B scale. R3-Reranker is the cross-encoder (rerank) stage of R3-Skill's two-stage retriever for query-conditional agent skill retrieval. It scores each (query, skill) pair jointly, paired with R3-Embedding-0.6B for recall.

View on Hugging Face → Fine-tuned from Qwen/Qwen3-Reranker-0.6B

Overview

Hardware: — drives latency, throughput & cost

Size	596M params
Tasks	/score
License	apache-2.0
Latency	449 ms
Throughput	2.1K tok/s
Cost	$0.103 /1M tok

Cost is approximate — computed from list GPU prices; your actual price depends on the provider you deploy SIE with.

Scoring

Inputs	text
Max sequence length	4,096

Benchmarks

R3Skill (candidates: R3-embedding-0.6b, k=20)

technology retrieval en

Agent skill retrieval: match user requests to the SKILL.md that solves them (Tencent R3 release set)

Quality

ndcg at 10 0.8228

Performance L4 b1 c16

Corpus 25.4K tok/s

Corpus p50 469.9ms

Query 2.2K tok/s

Query p50 469.9ms

Reference →

R3Skill (candidates: R3-embedding-0.6b, k=20, limit 384)

technology retrieval en

Agent skill retrieval: match user requests to the SKILL.md that solves them (Tencent R3 release set)

Quality

ndcg at 10 0.8264

Performance L4 b1 c16

Corpus 25.1K tok/s

Corpus p50 429.0ms

Query 2.1K tok/s

Query p50 429.0ms

Performance RTX-PRO-6000-Blackwell-Server-Edition b1 c16

Corpus 118.0K tok/s

Corpus p50 105.6ms

Query 9.8K tok/s

Query p50 105.6ms