What is FireRed-OCR?
FireRed-OCR is an open document-parsing model from Xiaohongshu's FireRedTeam, released on 28 February 2026 under Apache 2.0 for both code and weights, with a technical report on 2 March 2026. It reads a whole page image and writes Markdown, with tables and LaTeX formulas, in a single pass and without a separate layout model. The Hugging Face API counts 2.13 billion parameters; the model is Qwen3-VL-2B-Instruct with its architecture unchanged.
The work is in the training. A three-stage curriculum moves from detection and region-reading tasks, to fine-tuning on standardised Markdown, to a reinforcement-learning stage the team calls Format-Constrained GRPO, which rewards valid formula syntax, closed tables and properly nested structure. Hard training pages were relabelled with help from Gemini 3 Pro, according to the report.
Is FireRed-OCR worth it?
FireRed-OCR is the strongest single-pass parser in its makers' own comparison, and the comparison shows its limits. The team reports 92.94 on OmniDocBench v1.5, ahead of DeepSeek-OCR 2 at 91.09 and dots.ocr at 88.41, and 11.07 points above the untouched Qwen3-VL-2B it starts from. In the same table PaddleOCR-VL-1.5 scores 94.50 and GLM-OCR 94.60, both of which run a layout detector before reading. Every figure is self-reported, and we found no outside rerun.
The team's own robustness test is less flattering. On FireRedBench, its set of awkward real-world layouts, FireRed-OCR scores 74.62 against 76.47 for PaddleOCR-VL-1.5 and 79.68 for Gemini 3.0 Pro. Table structure on PubTabNet, 77.0, sits below MinerU2.5 at 88.4. The gains are real on formulas and text edit distance, 91.71 CDM and 0.032, but tables remain where a dedicated pipeline does better.
The practical appeal is that nothing about it is exotic. Because the architecture is Qwen3-VL, it loads through the standard Transformers class without trust_remote_code, and the BF16 weights are about 4.3 GB, small enough for one consumer GPU. The card documents only a Transformers example with an 8,192-token output cap and a prompt helper cloned from GitHub; it gives no vLLM recipe, VRAM figure or pages-per-second number.
Activity is the other worry. The GitHub repository, with 295 stars, has had no commits since 4 March 2026, and there has been one checkpoint. PaddleOCR-VL-1.6 is half the size and reports higher scores, and DeepSeek-OCR 2, which FireRed-OCR beats, at least has independent reruns to compare against.
Where it earns its place
An ML engineer whose stack already serves Qwen3-VL models
Add page-to-Markdown parsing without a second inference framework
The checkpoint uses the stock Qwen3VLForConditionalGeneration class, so it slots into existing Transformers code. Expect to write your own batching and serving, since the card publishes no throughput figures.
A researcher digitising maths-heavy lecture notes and papers
Get valid LaTeX out of scanned pages without broken formulas
The reinforcement-learning stage specifically rewards formula syntax, and the team reports 91.71 CDM on OmniDocBench v1.5. Check tables by hand, where pipeline parsers score higher.
An applied scientist studying OCR training methods
Reproduce format-constrained reinforcement learning on another base model
The technical report sets out the three-stage curriculum and the reward terms for formulas, tables, hierarchy and text, and Apache 2.0 weights give a 2B baseline to fine-tune from.
Before you commit
- Setup
- pip install transformers and qwen-vl-utils, clone the FireRed-OCR GitHub repository for its prompt helper, then load FireRedTeam/FireRed-OCR in BF16; flash-attention 2 is recommended.
- Integrations
- Transformers only on the card; outputs Markdown with tables and LaTeX formulas. A Hugging Face Space hosts a demo.
- Data and privacy
- Self-hosted, so documents stay on your hardware. Apache 2.0 permits commercial use and fine-tuning; the report says hard training samples were relabelled with Gemini 3 Pro.
- Cost at scale
- Weights are free and about 4.3 GB in BF16; no hosted price, VRAM figure or pages-per-second number is published, so fleet sizing needs your own benchmark.
- Limits
- Every score is self-reported; trails pipeline parsers on tables and on the team's own FireRedBench; output capped at 8,192 new tokens in the example; one checkpoint and no commits since 4 March 2026.
What’s good
- 92.94 on OmniDocBench v1.5, top among end-to-end models in its makers' table
- Standard Qwen3-VL architecture, no remote code to trust
- About 4.3 GB of BF16 weights
- Apache 2.0 for code and weights
What’s not
- All benchmark figures are self-reported
- Trails PaddleOCR-VL-1.5 and GLM-OCR on both of its own benchmarks
- No serving guide, VRAM or throughput figures
- No repository commits since March 2026
Best for
Teams already serving Qwen3-VL who want a drop-in document-to-Markdown checkpoint without adding a layout pipeline.
FireRed-OCR alternatives
FireRed-OCR: common questions
- Is FireRed-OCR worth it?
- A clear demonstration that reinforcement learning on format can lift a 2B general model by 11 points, but not yet a reason to leave the pipeline parsers that still beat it. We rate it 3.9 out of 5.
- How much does FireRed-OCR cost?
- Free under Apache 2.0. No hosted API; a free Hugging Face Space demo only.
- What are the downsides of FireRed-OCR?
- All benchmark figures are self-reported; trails PaddleOCR-VL-1.5 and GLM-OCR on both of its own benchmarks; no serving guide, VRAM or throughput figures; no repository commits since March 2026.
- Who is FireRed-OCR best for?
- Teams already serving Qwen3-VL who want a drop-in document-to-Markdown checkpoint without adding a layout pipeline.
- How much does FireRed-OCR cost for a team or in production?
- Weights are free and about 4.3 GB in BF16; no hosted price, VRAM figure or pages-per-second number is published, so fleet sizing needs your own benchmark.
- Is FireRed-OCR safe for company data?
- Self-hosted, so documents stay on your hardware. Apache 2.0 permits commercial use and fine-tuning; the report says hard training samples were relabelled with Gemini 3 Pro.
- What are the best FireRed-OCR alternatives?
- The closest alternatives we have reviewed are PaddleOCR-VL-1.6, DeepSeek-OCR 2, dots.ocr and Falcon OCR.
Answers checked Nov 2, 2026.