Turn scans, forms and tables into JSON

Turn scans, forms and tables into structured JSON.

74 models labelled for document OCR at these constraints. Data observed 2026-10-25 06:54 UTC.

Showing every model that fits the job; narrow it with the constraints below.

What it looks like in practice

About $0.15 to $0.75 per 1,000 pages across the 8 cheapest hosted models labelled for document OCR.

≈ $0.15 per 1,000 pages with DeepSeek-V4.1-Flash (assuming ~1,500 input and ~500 output tokens per page; your documents will differ).

Source: Cheapest live endpoint with both input and output prices, observed 2026-10-25

Start with these three

One pick per constraint group: cheapest per 1,000 pages hosted, most used that fits one 24 GB GPU, and the highest olmOCR-bench score (labelled when it is the author's own number).

Hosted, cheapest per 1,000 pages

DeepSeek-V4.1-Flash

deepseek-ai

Lowest estimated cost per 1,000 pages among hosted models labelled for document OCR: about $0.15 (Sail Research, 2026-10-24).

≈ $0.15per 1,000 pages

$0.020per 1M input

763Bparams

Sources: openrouter.ai 2026-10-24 · huggingface.co 2026-10-25

Self-host on one GPU

gemma-4-26B-A4B-it

google

Most downloaded in the last 30 days among document-OCR models whose INT4 weights fit one 24 GB GPU: 8.6M (Hugging Face, 2026-10-25).

~16 GBINT4 est.

8.6Mdownloads 30d

25.8Bparams

Sources: estimate method 2026-10-25 · huggingface.co 2026-10-25

Highest author-reported

chandra-ocr-2

datalab-to

Highest author-reported olmOCR-bench score among models labelled for document OCR: 85.8% (huggingface.co, 2026-10-07).

85.8%olmOCR-bench

5.30Bparams

~3.2 GBINT4 est.

Sources: huggingface.co 2026-10-07 · huggingface.co 2026-10-25 · estimate method 2026-10-25

Recipes

How people wire a model into this job. Notes are written by hand from the cited cards and docs; the model links go to the facts.

Retrieval stack with bge-m3 + a small MoE generator

Model: BAAI/bge-m3
Harness: Embedding endpoint (a serverless provider or self-hosted TEI) + any OpenAI-compatible chat endpoint
Weights: FP16 embeddings; FP8 generator

Chunk to the embedding model's max sequence length (card says 8192 tokens for bge-m3).
Use the generator's structured-output flag (provider row) for JSON extraction.
Re-embed on model change; keep the embedding model id in the index metadata.

Everything that fits 74

The first ten models labelled for document OCR, in the scoreboard's order for this job. The full list, sortable, lives in the scoreboard with these constraints.

Everything that fits: first 10 of 74
#
1	DeepSeek-V4.1-Flashdeepseek-ai	≈ $0.15	$0.020	34	~458 GB	1M	MIT
2	gemma-4-26B-A4B-itgoogle	≈ $0.17	$0.042	16	~16 GB	256K	Apache 2.0
3	gemma-4-26B-A4Bgoogle	≈ $0.17	$0.042	15	~16 GB	256K	Apache 2.0
4	gemma-4-31B-itgoogle	≈ $0.32	$0.10	14	~19 GB	256K	Apache 2.0
5	gemma-4-31Bgoogle	≈ $0.32	$0.10	13	~19 GB	256K	Apache 2.0
6	Qwen3-VL-30B-A3B-InstructQwen	≈ $0.53	$0.15	4	~19 GB	256K	Apache 2.0
7	Qwen3-VL-235B-A22B-InstructQwen	≈ $0.74	$0.20	3	~141 GB	256K	Apache 2.0
8	Qwen3-VL-8B-InstructQwen	≈ $0.75	$0.25	2	~5.3 GB	256K	Apache 2.0
9	GLM-4.6V-Flashzai-org	≈ $0.90	$0.30	2	~6.2 GB	128K	MIT
10	GLM-4.6Vzai-org	≈ $0.90	$0.30	2	~65 GB	128K	MIT

Show all 74 in the scoreboard →

Most-shared on this job

Links about these models ranked by how many newsletters and link sources shared them in the last 28 days; each source counts once.All links

Shared by 5
DeepSeek-V4.1-Flash-1017 (Hugging Face Repo)Hacker News, The Neuron, TLDR AI, AI Breakfast and 1 moreAbout DeepSeek-V4.1-Flash
huggingface.co · first seen 2026-10-17
Shared by 2
DeepSeek V4.1 Flash: Large-Scale Inference DeploymentLatent Space / AINews, TLDR AIAbout DeepSeek-V4.1-Flash
zhuanlan.zhihu.com · first seen 2026-10-21
Shared by 2
DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made RealThe Neuron, Turing PostAbout DeepSeek-V4.1-Flash
api-docs.deepseek.com · first seen 2026-10-19
Shared by 2
Chinese AI lab DeepSeek updated its V4.1-Flash model, which it says now finishes tasks faster with shorter answers, and…The Rundown AI, Don't Worry About the VaseAbout DeepSeek-V4.1-Flash
x.com · first seen 2026-10-19
Shared by 2
DeepSeek V4.1 Flash 1017Hacker News, AlphaSignalAbout DeepSeek-V4.1-Flash
openrouter.ai · first seen 2026-10-17

How people judge this job

None of them measure cost or field accuracy on your own documents. The cost line assumes about 1,500 input and 500 output tokens per page; dense forms, tables and long pages cost more.

Nearest scoreboard view: Cheapest hosted. Follow changes to this job's evidence: RSS.