What is Infinity-Parser2-Flash?
Infinity-Parser2-Flash is the smaller of two document parsers from INF Technology, a Shanghai AI company that publishes on Hugging Face as infly. Both were released on 11 May 2026 under Apache 2.0, with a technical report on arXiv. Flash is a fine-tuned Qwen3.5-2B, 2.21 billion parameters; its sibling Infinity-Parser2-Pro is built on Qwen3.5-35B-A3B, a 35-billion-parameter mixture-of-experts model that activates about 3 billion per token.
One model returns a page as JSON with a bounding box, category and text for each element, or as Markdown, and also parses charts to tables, chemical structures and document questions. INF trained it with supervised fine-tuning and then reinforcement learning across all these tasks at once.
INF also released Infinity-Doc2-5M, the training set of nearly 5 million synthetic and real document samples. Flash drew 15,305 downloads in the past month on Hugging Face; Pro, which needs two GPUs in INF's set-up, drew 2,334.
Is Infinity-Parser2-Flash worth it?
Infinity-Parser2-Flash is a capable all-rounder whose best numbers belong to its larger sibling. INF reports 86.0 on olmOCR-Bench and 72.2 on ParseBench for Flash, against 87.6 and 74.3 for the 35B Pro, and both well above the 78.5 and 66.0 it measured for PaddleOCR-VL-1.5. On OmniDocBench v1.6, though, Flash scores 91.98 in INF's table, below that same PaddleOCR-VL-1.5 at 94.87 and well short of the 96.47 OpenDataLab lists for OvisOCR2. Neither Infinity model appears on OpenDataLab's leaderboard, and INF marks several rival scores as run with its own tools.
Outside comparisons so far cover only Pro. LightOn's October benchmarks put Infinity-Parser2-Pro first on olmOCR-Bench at 87.6, matching INF's figure, but second on ParseBench at 74.3 behind LightOnOCR-3-4B's 75.1, and fifth of nine on LightOn's French-document test at 63.2. Those are a rival's tables with its own normalisation, but they are the only outside look at the family so far.
The speed claim needs reading closely. The headline 1,624 tokens a second, 3.68 times the older Infinity-Parser-7B, was measured on two H100s with eight requests in flight, which works out to 0.95 seconds a page. On one H100 at one request, the setting INF used for its rivals, Flash takes 6.06 seconds a page, against 2.48 for PaddleOCR-VL-1.5 and 2.35 for DeepSeek-OCR-2. Its JSON output, about 1,500 tokens a page, is part of the reason.
INF's own limitations section is candid: English and Chinese are the primary languages and others degrade, rotated tables and complex charts lose accuracy, bold and italic are not captured, and the model is weak at following multi-step instructions. Pro beats Flash on every benchmark INF lists except DocLayNet layout, but at 35.1 billion parameters it is a different purchase.
Our take: Flash makes sense where one model has to cover pages, charts and chemical structures, such as in pharmaceutical or research archives. For converting ordinary business documents to Markdown, it costs more GPU time per page than parsers that score higher.
Where it earns its place
A cheminformatics team at a pharmaceutical company digitising patents and lab reports
Turn chemical structure drawings and data tables into machine-readable text
Run the infinity_parser2 package over directories of PDFs on a vLLM backend. INF reports 63.34 on CoSyn chemical parsing for Flash and 73.19 for Pro; check a sample, since Pro's lead on this task is wide.
A data engineer at an asset manager parsing quarterly reports
Extract chart data and tables into JSON with positions for later review
Use the doc2json task to get each element with its box, then feed tables and chart tables to a database. Batch eight pages at once over two GPUs to get near INF's 0.95 seconds a page.
Before you commit
- Setup
- Install torch 2.10.0, flash-attn 2.8.3 (10 to 30 minutes to compile from source) and vllm 0.17.1, then pip install infinity_parser2 and run the parser command on a file or a whole directory; weights download on first run.
- Integrations
- Python API and CLI with vLLM engine, vLLM server and Transformers backends; doc2json, doc2md and custom-prompt tasks; community GGUF, AWQ and MLX conversions on Hugging Face.
- Data and privacy
- Apache 2.0 for weights; the Infinity-Doc2-5M dataset is public. Self-hosted only, so documents stay on your hardware.
- Cost at scale
- The BF16 checkpoint is about 4.4 GB. INF's figures are 0.95 seconds a page on two H100s at concurrency 8, about 3,800 pages an hour for the pair, and 6.06 seconds a page single-stream.
- Limits
- Primarily English and Chinese; rotated tables, complex charts and text styling are weak points; unreliable on complex multi-step instructions, per INF.
What’s good
- 86.0 on olmOCR-Bench and 72.2 on ParseBench (self-reported)
- Charts, chemical structures and document questions in one 2.2B model
- JSON output with per-element boxes and categories
- Apache 2.0, with the 5-million-sample training set released
What’s not
- 91.98 on OmniDocBench v1.6 by INF's count; not on the public leaderboard
- 6.06 seconds a page on one H100 without batching
- English and Chinese only in practice
- No first-party API; heavy install with pinned torch, vLLM and FlashAttention
Best for
Research and technical-document teams who want page parsing, chart extraction and chemical-structure recognition from one small self-hosted model.
Infinity-Parser2-Flash alternatives
Infinity-Parser2-Flash: common questions
- Is Infinity-Parser2-Flash worth it?
- Infinity-Parser2-Flash does more kinds of document work than most 2B models, but on plain page parsing it trails smaller parsers and needs batching to be fast. We rate it 3.8 out of 5.
- How much does Infinity-Parser2-Flash cost?
- Free under Apache 2.0. No first-party hosted API; a free demo runs on Hugging Face Spaces.
- What are the downsides of Infinity-Parser2-Flash?
- 91.98 on OmniDocBench v1.6 by INF's count; not on the public leaderboard; 6.06 seconds a page on one H100 without batching; english and Chinese only in practice; no first-party API; heavy install with pinned torch, vLLM and FlashAttention.
- Who is Infinity-Parser2-Flash best for?
- Research and technical-document teams who want page parsing, chart extraction and chemical-structure recognition from one small self-hosted model.
- How much does Infinity-Parser2-Flash cost for a team or in production?
- The BF16 checkpoint is about 4.4 GB. INF's figures are 0.95 seconds a page on two H100s at concurrency 8, about 3,800 pages an hour for the pair, and 6.06 seconds a page single-stream.
- Is Infinity-Parser2-Flash safe for company data?
- Apache 2.0 for weights; the Infinity-Doc2-5M dataset is public. Self-hosted only, so documents stay on your hardware.
- What are the best Infinity-Parser2-Flash alternatives?
- The closest alternatives we have reviewed are LightOnOCR-3-4B, PaddleOCR-VL-1.6, Chandra OCR 2 and Qianfan-OCR.
Answers checked Nov 2, 2026.