olmOCR-2-7B-1025

8.29B parameters, images and text in, text out, 125K tokens of context.

Verified 2026-10-25 Rank #664 in LLMs · score 72.9 method v1 Sources 3 Downloads 30d 76.8k · #344 of 619 tracked

Can I run it?5.0GB at INT4Estimatefits a 16 GB cardmethod

Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24

Do people use it?76.8kdownloads, 30 days#344 of 619 we track · below the median+0 likes, downloads −4% this week159 likesHugging Face2026-10-24

Worth reading

Picked by rule from 2 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Latest · 12 months ago

OlmOCR 2: Unit test rewards for document OCR

AllenAI releases olmOCR 2, a vision-language model built on Qwen2.5-VL-7B that achieves 82.4 points on olmOCR-Bench for English digitized print documents.

The synthetic data pipeline costs $0.12 per page to generate 2,186 PDF pages with 30,381 verifiable test cases.
Training generates 28 completions per document, rewarding only those that pass more unit tests.

allenai.org 2025-10-22 3 HN pts · 1 comments 876 words about the olmOCR 2 family family

2 best of 2 tracked

The 1 pick above, plus these 1. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

CodeOlmOCR 2github.com · 2025-10-24 · 5 HN pts

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 11 quantized builds and 17 Spaces.

As a service0no API provider listed
On own hardware11quantized builds
Built on166 fine-tunes · 10 adapters · 0 merges
In apps17Hugging Face Spaces

Top Spaces running it: prithivMLmods/Multimodal-OCR3 ♥ 69 · hfmlsoc/different-flops ♥ 11 · IFMedTechdemo/Multi-Model-OCR ♥ 7

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 5.0 GB at INT4, 10 GB at FP8, 20 GB at BF16: weights only, so leave room for the KV cache.

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

2 dated articles and 9 repo commits so far

Teal is the weights (2025-10-06), amber triangles are 2 dated articles (each links to it), grey ticks are repo commits.

What people built on it

27 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

27 derivatives on the Hub, led by 11 quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF lmstudio-community/olmOCR-2-7B-1025-GGUF for llama.cpp, Ollama, LM Studio 210k downloads MLX richardyoung/olmOCR-2-7B-1025-MLX-8bit for Apple silicon 754 downloads INT4 winninghealth/olmOCR-2-7B-1025-INT4 for vLLM, SGLang 249.8k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

18 rows

Benchmarks 18 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (18)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
author-reported olmOCR-bench · Extract text from documents	82.9	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	82.1	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	84.3	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	48.3	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	95.7	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	84.3	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	81.4	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	99.7	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	82.3 ± 1.1	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	83	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025-FP8	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	82.3	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025-FP8	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	84.9	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025-FP8	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	47.7	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025-FP8	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	96.1	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025-FP8	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	83.7	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025-FP8	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	81.9	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025-FP8	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	99.7	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025-FP8	model author	2026-10-07
author-reported olmOCR-bench · Extract text from documents	82.4 ± 1.1	olmOCR pipeline v0.4.0 with olmOCR-2-7B-1025-FP8	model author	2026-10-07

Use-case evidence

4 use cases

What it is good at 4 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

20 fields

Listed in LLMs

Facts

olmOCR-2-7B-1025 has 8.29B parameters. (safetensors metadata, observed 2026-10-24)
Context length is 128,000 tokens (125K). (config.json max_position_embeddings, observed 2026-10-24)
License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-24)
Created 2025-10-06, last modified 2025-10-22. (HF API, observed 2026-10-24)

Pipeline tag	image-text-to-text
Library	transformers
Architectures	Qwen2_5_VLForConditionalGeneration
Model type	qwen2_5_vl
Dtype	bfloat16
Params by dtype	BF16: 8,292,166,656
Size band	M
MoE	no
Context length	128,000 (config.json max_position_embeddings)
Languages	en
License	Apache 2.0 (apache-2.0)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Qwen2_5_VLForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Qwen2_5_VLForConditionalGeneration · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for olmOCR-2-7B-1025: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	20 GB · Single GPU (<24 GB)
FP8 (1 byte/param)	10 GB · Single GPU (<24 GB)
INT4 (0.5 bytes/param)	5.0 GB · Single GPU (<24 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

27 variants

Lineage

Fine-tuned from Qwen/Qwen2.5-VL-7B-Instruct Family: allenai/olmOCR-2 27 derivatives on the Hub

Base model: Qwen/Qwen2.5-VL-7B-Instruct

quantized: 11finetune: 6adapter: 10

Top variants (27 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
winninghealth/olmOCR-2-7B-1025-INT4	quantized	INT4	1.77B	S	apache-2.0	winninghealth	1	249.8k
lmstudio-community/olmOCR-2-7B-1025-GGUF	quantized	GGUF	—	—	apache-2.0	lmstudio-community	5	210k
bartowski/allenai_olmOCR-2-7B-1025-GGUF	quantized	GGUF	—	—	—	bartowski	5	30.9k
Charitarth/olmOCR-2-7B-1025-GPTQ-W4A16	quantized	GPTQ	2.63B	M	apache-2.0	Charitarth	2	14k
mradermacher/olmOCR-2-7B-1025-i1-GGUF	quantized	GGUF	—	—	apache-2.0	mradermacher	1	8.3k
mradermacher/olmOCR-2-7B-1025-GGUF	quantized	GGUF	—	—	apache-2.0	mradermacher	2	7.5k
richardyoung/olmOCR-2-7B-1025-GGUF	quantized	GGUF	—	—	apache-2.0	richardyoung	4	3.5k
orgessnk/olmOCR-2-7B-1025-BNB-4bit	quantized	-BNB	8.29B	M	apache-2.0	orgessnk	1	1.2k
Charitarth/olmOCR-2-7B-1025-NVFP4	quantized	NVFP4	5.44B	M	apache-2.0	Charitarth	0	1.1k
richardyoung/olmOCR-2-7B-1025-MLX-8bit	finetune	MLX	8.29B	M	apache-2.0	richardyoung	0	754
richardyoung/olmOCR-2-7B-1025-MLX-6bit	finetune	MLX	8.29B	M	apache-2.0	richardyoung	0	444
richardyoung/olmOCR-2-7B-1025-MLX-4bit	finetune	MLX	8.29B	M	apache-2.0	richardyoung	0	392
cstr/olmOCR-2-7B-1025-GGUF	quantized	GGUF	—	—	apache-2.0	cstr	0	384
Naruto123321/olmOCR-2-7B-1025-Q2_K-GGUF	quantized	Q2_K	—	—	apache-2.0	Naruto123321	0	141
hastyle/olmOCR-arabic-lora	adapter	—	—	—	apache-2.0	hastyle	0	118
suv11235/olmOCR-7B-grpo-v3	finetune	—	8.29B	M	—	suv11235	1	57
hastyle/olmOCR-arabic-lora-v2	adapter	—	—	—	apache-2.0	hastyle	0	81
enclavelabs/enclave-scribe-devanagari	adapter	—	—	—	mit	enclavelabs	0	65
suv11235/chandra-grpo-lora-10-ep-50-docs	adapter	—	—	—	apache-2.0	suv11235	0	58
suv11235/olmOCR-7B-grpo-v6	finetune	—	8.29B	M	—	suv11235	0	45
suv11235/olmOCR-7B-grpo-v2	finetune	—	8.29B	M	—	suv11235	0	41
suv11235/chandra-grpo-5ep-4doc-full-extraction	adapter	—	8.29B	M	—	suv11235	0	37
enclavelabs/enclave-scribe-devanagari-iter4	adapter	—	—	—	mit	enclavelabs	0	36
rehan953/olmOCR-2-7B-1025-finetuned_v2	adapter	—	—	—	apache-2.0	rehan953	0	34
rehan953/olmOCR-2-7B-1025-finetuned	adapter	—	—	—	apache-2.0	rehan953	0	33
enclavelabs/olmocr-2-iter7a-vqa	adapter	—	—	—	mit	enclavelabs	0	24
larshiakzemil/olmocr-farsi-lora	adapter	—	—	—	apache-2.0	larshiakzemil	0	22

Repository changelog

9 commits

Repository changelog 9 commits

Date	Commit	SHA
2025-10-22	Update README.md	e52d6f0
2025-10-22	Update README.md	30cc296
2025-10-21	Update README.md	a124f0c
2025-10-17	Update README.md	b394d20
2025-10-06	Update README.md	a38c129
2025-10-06	Update README.md	24f9c62
2025-10-06	Update README.md	84424bf
2025-10-06	Upload folder using huggingface_hub	24a2443
2025-10-06	initial commit	937861f

Articles and coverage

2 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official1

OlmOCR 2: Unit test rewards for document OCR

allenai.org2025-10-22officialauthorityabout the olmOCR 2 familyHN 3 pts · 1 comments876 words

AllenAI releases olmOCR 2, a vision-language model built on Qwen2.5-VL-7B that achieves 82.4 points on olmOCR-Bench for English digitized print documents. The system fine-tunes on a dataset of 270,000 PDF pages and uses Group Relative Policy Optimization with unit test rewards to generate structured Markdown, HTML, and LaTeX in a single pass. An FP8 quantized version delivers 3,400 output tokens/sec on a single H100 GPU, enabling processing of 10,000 pages for less than $2.written, reviewed, cited

The synthetic data pipeline costs $0.12 per page to generate 2,186 PDF pages with 30,381 verifiable test cases.
Training generates 28 completions per document, rewarding only those that pass more unit tests.
The model correctly interprets the date in Abraham Lincoln's January 10, 1864 letter to Major General Hitchcock.

Analysis1

OlmOCR 2

github.com2025-10-24authorityabout the olmOCR 2 familyHN 5 pts · 1 comments2,075 words

olmOCR converts PDFs and image-based documents into clean Markdown using a 7B parameter VLM that requires a GPU. The October 21, 2025 v0.4.0 release boosts olmOCR-bench scores by ~4 points using synthetic data and introduces RL training. Conversion costs less than $200 USD per million pages.written, reviewed, cited

Local GPU inference requires a recent NVIDIA GPU with at least 12 GB of GPU RAM and 30GB of free disk space.
The v0.2.1 release runs significantly faster because it defaults to FP8 precision and needs much fewer retries per document.
Remote inference on Cirrascale costs $0.07 per million input tokens and $0.15 per million output tokens.

Similar models

3 models

Similar models

Qwen3-VL-8B-Instruct · 8.77B
same size band (M), same task (image-text-to-text), more providers, permissive license
Qwen2.5-VL-7B-Instruct · 8.29B
same size band (M), same task (image-text-to-text), permissive license
Qwen3.5-9B · 9.65B
same task (image-text-to-text), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:41 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:13 UTC · Last verified 2026-10-25

Follow olmOCR-2-7B-1025

Watch this model

/feeds/models/allenai--olmOCR-2-7B-1025.xml(RSS 2.0) · every change to olmOCR-2-7B-1025
/feeds/families/allenai--olmOCR-2.xml(RSS 2.0) · every change across the olmOCR-2 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.