OCR models / FireRed team (Xiaohongshu)

FireRed-OCR

Name: FireRed-OCR
Author: FireRed team (Xiaohongshu)

FireRed team (Xiaohongshu) · released 2026-02-28 · data checked 2026-11-06

FireRed-OCR is a 2.1B-parameter OCR and document-parsing model from FireRed team (Xiaohongshu) (CN), released 2026-02-28 under the Apache-2.0 license. It scores 93.26 on OmniDocBench v1.6 (#20 of 33) and 70.2 on olmOCR-Bench (#22 of 24). It needs about 5.3 GB of memory at 16-bit.

Built on Qwen3-VL 2B.

OmniDocBench v1.6

93.26

#20 of 33 models

olmOCR-Bench

70.2

#22 of 24 models

Parameters

2.1B

OCR specialist

Memory to run

~5.3 GB

16-bit weights plus about 1 GB

Benchmarks

OmniDocBench v1.6 Leaderboard93.26

Best: OXR 1.0 97.03

olmOCR-Bench Leaderboard70.2

Headers and footers excluded. Best: OXR 1.0 87.8

The tick marks the best score on each benchmark. See how to read the ranking.

Specifications

Lab: FireRed team (Xiaohongshu) (CN)
Type: OCR / document parsing specialist
Total parameters: 2.1B
License: Apache-2.0
Memory to run: ~5.3 GB at 16-bit
Released: 2026-02-28

Weights and usage on Hugging Face ↗

Closest alternatives

ABot-OCR (2.1B, 93.30 on OmniDocBench, Not stated)
Logics-Parsing v2 (4.4B, 93.33 on OmniDocBench, Not stated)
MinerU2.5 (2509) (1.2B, 93.04 on OmniDocBench, AGPL-3.0)
Ovis2.6 30B-A3B (31.4B, 93.70 on OmniDocBench, Apache-2.0)

Sources

Something wrong here? Report it.