Model Follower.

Up to date

Rebuilt
2026-10-26 07:36 UTC
Release radar
116 on watch
Hub snapshot
2026-10-26
Re-checked
daily · every number sourced and dated

Models › Unlimited family

Unlimited-OCR

Hugging Face ↗ Follow changes

baidu MIT out 4 months ago

3.34B parameters (6 of 64 experts per token), images and text in, text out, 32K tokens of context.

Verified 2026-10-26 07:27 UTC Rank #72 in LLMs · score 104 method v1 Sources 3 Downloads 30d 2.6M · #55 of 619 tracked

Changed this week: 2 new quantized builds on the Hub (39 in all).2026-10-24 · 1 new quantized build on the Hub (38 in all).2026-10-24

Can I run it?2.0GB at INT4Estimatefits a 16 GB cardmethod
Can I use it?MITPermissivelicence file2026-10-26
Do people use it?2.6Mdownloads, 30 days#55 of 619 we track · top 9%+119 likes, downloads +56% this week4,892 likesHugging Face2026-10-26

Worth reading

Picked by rule from 12 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 496 HN pts

Unlimited OCR: One-shot long-horizon parsing

Baidu released Unlimited-OCR, a model designed to advance document parsing capabilities beyond previous solutions.

  • Single image inference supports two distinct configurations: gundam and base.
  • Multi-page PDF parsing exclusively uses the base configuration with an image size of 1024.
github.com 2026-06-23 496 HN pts · 110 comments 901 words
Latest · 4 months ago

Unlimited OCR Works

Unlimited OCR replaces all attention layers in the decoder with Reference Sliding Window Attention to maintain a constant KV cache throughout decoding.

arxiv.org 2026-06-22 3 HN pts 396 words

12 best of 12 tracked

The 3 picks above, plus these 9. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. PaperUnlimiformer: Long-Range Transformers with Unlimited Length Inputarxiv.org · 2023-05-05 · 335 HN pts
  2. Paper[PDF] Unlimited OCR Works - arXivarxiv.org · undated
  3. ArticlePaper page - Unlimited OCR Works - Hugging Facehuggingface.co · undated
  4. ArticleUnlimited OCR: What Baidu's 3B Model Can and Can't Doaireiter.com · 2026-07-28
  5. ArticleBaidu's Unlimited-OCR Reads a 40-Page PDF in One Pass — Here's What the Viral Posts Get Wrong, and What Actually Matters - Thorsten Meyer AIthorstenmeyerai.com · 2026-07-21
  6. ArticleUnlimited-OCR Just Ripped Up the Rulebook on Document Parsing - Banandrebanandre.com · 2026-06-25
  7. ArticleBaidu Unlimited-OCR: one long document | iiterateiiterate.de · 2026-06-25
  8. Article[Literature Review] Unlimited OCR Worksthemoonlight.io · undated
  9. ArticleUnlimited-OCR Explained: How Baidu Solved Long-Document AIanalyticsvidhya.com · undated

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: 39 quantized builds and 78 Spaces.

  • As a service0no API provider listed
  • On own hardware39quantized builds
  • Built on1916 fine-tunes · 3 adapters · 0 merges
  • In apps78Hugging Face Spaces

Top Spaces running it: baidu/Unlimited-OCR ♥ 495 · embedl/hfviewer ♥ 181 · akhaliq/Unlimited-OCR ♥ 52 · prithivMLmods/DeepSeek-OCR-2-Unlimited-OCR ♥ 25

Will it fit?

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

Fits a 16 GB card at INT4

About 2.0 GB at INT4, 4.0 GB at FP8, 8.0 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 2.0 GB of weights (estimate; KV cache and activations extra)INT42.0 GBFP8: about 4.0 GB of weights (estimate; KV cache and activations extra)FP84.0 GBBF16: about 8.0 GB of weights (estimate; KV cache and activations extra)BF168.0 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 2.0 GB of weights (estimate; KV cache and activations extra)INT42.0 GBFP8: about 4.0 GB of weights (estimate; KV cache and activations extra)FP84.0 GBBF16: about 8.0 GB of weights (estimate; KV cache and activations extra)BF168.0 GB

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

The story so far

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

Most coverage came in the first two weeks

Teal is the weights (2026-06-19), amber triangles are 8 dated articles (each links to it), grey ticks are repo commits.

JunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2026-10-14: 1 repo commit2026-07-29: 1 repo commit2026-07-28: 1 repo commit2026-07-23: 1 repo commit2026-07-21: 1 repo commit2026-07-03: 1 repo commit2026-06-28: 2 repo commits2026-06-24: 1 repo commit2026-06-23: 2 repo commits2026-06-22: 2 repo commitsUnlimited OCR: One-shot long-horizon parsing · 2026-06-23Unlimited OCR Works · 2026-06-22Unlimiformer: Long-Range Transformers with Unlimited Length Input · 2023-05-05Unlimited-OCR on Real Documents: Local Results | VoicePing · 2026-07-30Unlimited OCR: What Baidu's 3B Model Can and Can't Do · 2026-07-28Baidu's Unlimited-OCR Reads a 40-Page PDF in One Pass — Here's What the Viral Posts Get Wrong, and What Actually Matters - Thorsten Meyer AI · 2026-07-21Unlimited-OCR Just Ripped Up the Rulebook on Document Parsing - Banandre · 2026-06-25Baidu Unlimited-OCR: one long document | iiterate · 2026-06-25Weights 2026-06-19 · 2026-06-19Weights 2026-06-19now
JunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2026-10-14: 1 repo commit2026-07-29: 1 repo commit2026-07-28: 1 repo commit2026-07-23: 1 repo commit2026-07-21: 1 repo commit2026-07-03: 1 repo commit2026-06-28: 2 repo commits2026-06-24: 1 repo commit2026-06-23: 2 repo commits2026-06-22: 2 repo commitsUnlimited OCR: One-shot long-horizon parsing · 2026-06-23Unlimited OCR Works · 2026-06-22Unlimiformer: Long-Range Transformers with Unlimited Length Input · 2023-05-05Unlimited-OCR on Real Documents: Local Results | VoicePing · 2026-07-30Unlimited OCR: What Baidu's 3B Model Can and Can't Do · 2026-07-28Baidu's Unlimited-OCR Reads a 40-Page PDF in One Pass — Here's What the Viral Posts Get Wrong, and What Actually Matters - Thorsten Meyer AI · 2026-07-21Unlimited-OCR Just Ripped Up the Rulebook on Document Parsing - Banandre · 2026-06-25Baidu Unlimited-OCR: one long document | iiterate · 2026-06-25Weights 2026-06-19 · 2026-06-19Weights 2026-06-19now

What people built on it

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

58 derivatives on the Hub, most of them quantized

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

39 quantized (Hugging Face model tree)16 fine-tunes (Hugging Face model tree)3 adapters (Hugging Face model tree)39 quantized16 fine-tunes3 adapters
39 quantized (Hugging Face model tree)16 fine-tunes (Hugging Face model tree)3 adapters (Hugging Face model tree)39 quantized16 fine-tunes3 adapters
GGUF sahilchachra/Unlimited-OCR-GGUF for llama.cpp, Ollama, LM Studio 140.5k downloads MLX sahilchachra/unlimited-ocr-mxfp8-mlx for Apple silicon 4.4k downloads AWQ sahilchachra/Unlimited-OCR-AWQ for vLLM, SGLang 3.5M downloads

All variants →

Reference

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

Use-case evidence

6 use cases

What it is good at 6 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

24 fields

Listed in LLMs · MoE

Facts

Pipeline tagimage-text-to-text
Librarytransformers
ArchitecturesUnlimitedOCRForCausalLM
Model typeunlimited-ocr
Dtypebfloat16
Params by dtypeBF16: 3,336,106,240
Size bandM
MoEyes
Experts64
Experts per token6
Context length32,768 (config.json max_position_embeddings)
Languagesmultilingual
PapersarXiv:2606.23050
LicenseMIT (mit)
Hub snapshot2026-10-26

Servability

vLLM · SGLang

Servability

vLLM: supported · UnlimitedOCRForCausalLM · 4b1c04e76fe9ac3ca8b7e541dd423e7d5f38dc60SGLang: supported · UnlimitedOCRForCausalLM · 60d941293e3af27a12c6eb64e2c3e2e1a17f100f

vLLM registry (4b1c04e76fe9ac3ca8b7e541dd423e7d5f38dc60) · SGLang registry (60d941293e3af27a12c6eb64e2c3e2e1a17f100f) · 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 Unlimited-OCR: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)8.0 GB · Single GPU (<24 GB)
FP8 (1 byte/param)4.0 GB · Single GPU (<24 GB)
INT4 (0.5 bytes/param)2.0 GB · Edge (<4 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

All variants

40 variants

Lineage

Original model Family: baidu/Unlimited 58 derivatives on the Hub

quantized: 39finetune: 16adapter: 3

Top variants (40 shown, ranked by downloads)

Relation
Quant
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
sahilchachra/Unlimited-OCR-AWQ quantized AWQ 3.36B M mit sahilchachra 16 3.5M
sahilchachra/Unlimited-OCR-GGUF quantized GGUF — — mit sahilchachra 155 140.5k
DevQuasar/baidu.Unlimited-OCR-GGUF quantized GGUF — — — DevQuasar 12 22.6k
sabafallah/Unlimited-OCR-GGUF quantized GGUF — — — sabafallah 5 15.3k
vimalnakrani/unlimited-ocr-gguf quantized gguf — — mit vimalnakrani 1 14.8k
sahilchachra/Unlimited-OCR-NVFP4 quantized NVFP4 2.20B M mit sahilchachra 3 11.5k
shadowrock-io/Unlimited-OCR-Community-FP8-Dynamic quantized -FP8 3.34B M mit shadowrock-io 1 9.3k
sahilchachra/unlimited-ocr-mxfp8-mlx quantized fp8- 3.34B M mit sahilchachra 15 4.4k
sabafallah/Unlimited-OCR-Universal finetune — 3.34B M mit sabafallah 7 3k
sahilchachra/unlimited-ocr-8bit-mlx quantized -8bit 3.34B M mit sahilchachra 4 2k
FluidInference/unlimited-ocr-mlx-coreml quantized mlx — — mit FluidInference 25 868
mlx-community/Unlimited-OCR-mxfp8 quantized — 3.34B M mit mlx-community 5 1.8k
schift-io/schift-ocr-1-beta finetune — 3.64B M mit schift-io 18 799
mlx-community/Unlimited-OCR-8bit quantized -8bit 3.34B M mit mlx-community 3 1.4k
sahilchachra/unlimited-ocr-4bit-mlx quantized -4bit 3.34B M mit sahilchachra 2 1.4k
mlx-community/Unlimited-OCR-bf16 finetune -bf16 3.34B M mit mlx-community 2 1.1k
mlx-community/Unlimited-OCR-4bit quantized -4bit 3.34B M mit mlx-community 2 1k
batiai/batisee quantized GGUF — — mit batiai 0 788
sahilchachra/unlimited-ocr-mxfp4-mlx quantized mxfp4 3.34B M mit sahilchachra 0 774
AyoubChLin/unlimited-ocr-quran-uthmani adapter — 3.34B M mit AyoubChLin 6 472
mikoy92/Unlimited-OCR-bf16-mlx finetune -bf16 3.34B M mit mikoy92 0 668
vimalnakrani/unlimited-ocr-bf16-mlx finetune -bf16 3.34B M mit vimalnakrani 1 412
majentik/Unlimited-OCR-MLX-8bit quantized MLX 3.34B M mit majentik 0 439
shadowrock-io/Unlimited-OCR-Community-NVFP4 quantized NVFP4 3.34B M mit shadowrock-io 1 387
mlboydaisuke/Unlimited-OCR-CoreAI quantized — — — mit mlboydaisuke 6 131
srv-sngh/Unlimited-OCR-mixed-nvfp4 finetune nvfp4 3.34B M mit srv-sngh 0 374
richardyoung/Unlimited-OCR-GGUF quantized GGUF — — mit richardyoung 0 374
majentik/Unlimited-OCR-MLX-6bit quantized MLX 3.34B M mit majentik 0 329
suryatmodulus/unlimited-ocr-mlx-coreml quantized mlx — — mit suryatmodulus 1 266
mlx-community/Unlimited-OCR-6bit quantized — 3.34B M mit mlx-community 0 311
pcuenq/Unlimited-OCR-bf16-mlx finetune -bf16 3.34B M mit pcuenq 1 259
vimalnakrani/unlimited-ocr-6bit-mlx quantized mlx 3.34B M mit vimalnakrani 0 293
corsairnui/Unlimited-OCR-AMD-Strix-Halo finetune — 3.34B M mit corsairnui 3 127
wittjeff/unlimited-ocr-6bit-mlx quantized mlx 3.34B M mit wittjeff 0 264
AutomatosX/AX-Unlimited-OCR-3B-MoE-MLX-AXQ-MXFP8 quantized MLX 3.34B M mit AutomatosX 0 245
vimalnakrani/unlimited-ocr-4bit-mixed-mlx quantized -4bit 3.34B M mit vimalnakrani 0 240
will702/unlimited-ocr-mxfp8-mlx quantized fp8- 3.34B M mit will702 0 218
emee-ai/Unlimited-OCR-bf16-mlx finetune -bf16 3.34B M mit emee-ai 0 179
vimalnakrani/unlimited-ocr-8bit-mlx quantized -8bit 3.34B M mit vimalnakrani 0 168
MUFFDUNCAN/Unlimited-OCR-Universal finetune — 3.34B M mit MUFFDUNCAN 1 103

Repository changelog

13 commits
changed in the last 7 daysOther: 2026-10-24, 2026-10-24all changes

Repository changelog 13 commits

DateCommitSHA
2026-10-14 Update files for compatibility with transformers (#13) 8249d20
2026-07-29 Update README.md 07dea83
2026-07-28 Update README.md 3f2e9c9
2026-07-23 Update README.md 27a5997
2026-07-21 Update README.md 2a06ebf
2026-07-03 Update README.md ee63731
2026-06-28 Update README.md 84757cb
2026-06-28 Update README.md dac4e40
2026-06-24 Update README.md 3a7f4db
2026-06-23 Update README.md f799a9c
2026-06-23 Update README.md f8ca26a
2026-06-22 Update README.md b1dcd43
2026-06-22 Initial commit d549bb9

Articles and coverage

12 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.

Analysis10

Unlimited OCR: One-shot long-horizon parsing
github.com2026-06-23authorityHN front pageHN 496 pts · 110 comments901 words
Baidu released Unlimited-OCR, a model designed to advance document parsing capabilities beyond previous solutions. The project supports vLLM inference and training with ms-swift, with the paper available on arXiv under identifier 2606.23050. It offers a maximum context-length of 32768 tokens and provides specific configurations for handling single images and multi-page PDFs.written, reviewed, cited
  • Single image inference supports two distinct configurations: gundam and base.
  • Multi-page PDF parsing exclusively uses the base configuration with an image size of 1024.
  • Deployment requires pinning kernels to version 0.11.7 when using the SGLang server.
Unlimited OCR Works
arxiv.org2026-06-22authorityHN 3 pts · 0 comments396 words
Unlimited OCR replaces all attention layers in the decoder with Reference Sliding Window Attention to maintain a constant KV cache throughout decoding. Building on the high compression rate of DeepSeek OCR's encoder, the model transcribes dozens of pages in a single forward pass under a standard maximum length of 32K. R-SWA serves as a general-purpose parsing attention mechanism applicable to tasks such as ASR and translation.written, reviewed, cited
Unlimiformer: Long-Range Transformers with Unlimited Length Input
arxiv.org2023-05-05authorityHN front pageabout the Unlimited familyHN 335 pts · 101 comments404 words
Unlimiformer wraps any pretrained encoder-decoder transformer to process practically unlimited input sequences by offloading cross-attention to a k-nearest-neighbor index. The method enables processing 500k token-long inputs from the BookSum dataset without input truncation at test time. It extends models like BART and Longformer to unlimited inputs without additional learned weights or code modifications.written, reviewed, cited
[PDF] Unlimited OCR Works - arXiv
arxiv.orgauthority6,302 words
Unlimited OCR from Baidu Inc. replaces decoder attention layers with Reference Sliding Window Attention to maintain a constant KV cache, enabling single-pass transcription of dozens of pages under a 32K maximum length. By attending to all reference tokens and the preceding 128 output tokens, the model achieves 93% accuracy on the OmniDocBench v1.5 benchmark, outperforming the DeepSeek OCR baseline by 6% while avoiding progressive memory growth during long-horizon decoding.written, reviewed, cited
  • DeepEncoder achieves a 16x token compression rate by cascading window attention ViT and global attention.
  • The decoder utilizes an MoE architecture with an activation size of only 500M during inference.
Paper page - Unlimited OCR Works - Hugging Face
huggingface.coauthority313 words
Unlimited OCR replaces all attention layers in the decoder of DeepSeek OCR with Reference Sliding Window Attention to maintain a constant KV cache. This mechanism allows the model to transcribe dozens of pages under a standard maximum length of 32K in a single forward pass. Beyond OCR, the general-purpose parsing attention mechanism is also applicable to tasks such as ASR and translation.written, reviewed, cited
Unlimited-OCR on Real Documents: Local Results | VoicePing
voiceping.net2026-07-301,984 words
Unlimited OCR: What Baidu's 3B Model Can and Can't Do
aireiter.com2026-07-282,109 words
Baidu's Unlimited-OCR Reads a 40-Page PDF in One Pass — Here's What the Viral Posts Get Wrong, and What Actually Matters - Thorsten Meyer AI
thorstenmeyerai.com2026-07-213,015 words
Unlimited-OCR Just Ripped Up the Rulebook on Document Parsing - Banandre
banandre.com2026-06-251,724 words
Baidu Unlimited-OCR: one long document | iiterate
iiterate.de2026-06-251,041 words
All coverage tracked for this model (12) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
[PDF] Unlimited OCR Works - arXivarxiv.org—2026-10-07exa
Paper page - Unlimited OCR Works - Hugging Facehuggingface.co—2026-10-07exa
[Literature Review] Unlimited OCR Worksthemoonlight.io—2026-10-07exa
Unlimited-OCR Explained: How Baidu Solved Long-Document AIanalyticsvidhya.com—2026-10-07exa
Unlimited-OCR on Real Documents: Local Results | VoicePingvoiceping.net2026-07-302026-10-07exa
Unlimited OCR: What Baidu's 3B Model Can and Can't Doaireiter.com2026-07-282026-10-07exa
Baidu's Unlimited-OCR Reads a 40-Page PDF in One Pass — Here's What the Viral Posts Get Wrong, and What Actually Matters - Thorsten Meyer AIthorstenmeyerai.com2026-07-212026-10-07exa
Unlimited-OCR Just Ripped Up the Rulebook on Document Parsing - Banandrebanandre.com2026-06-252026-10-07exa
Baidu Unlimited-OCR: one long document | iiterateiiterate.de2026-06-252026-10-07exa
Unlimited OCR: One-shot long-horizon parsinggithub.com2026-06-232026-10-07hn
Unlimited OCR Worksarxiv.org2026-06-222026-10-07hn, exa
Unlimiformer: Long-Range Transformers with Unlimited Length Inputarxiv.org2023-05-052026-10-07hn

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

Similar models

3 models

Similar models

  • DeepSeek-OCR · 3.34B
    same size band (M), also MoE, same task (image-text-to-text), more providers, permissive license
  • DeepSeek-OCR-2 · 3.39B
    same size band (M), also MoE, same task (image-text-to-text), permissive license
  • deepseek-vl2-tiny · 3.37B
    same size band (M), same task (image-text-to-text)

Sources

3 sources

Sources

Hub snapshot: 2026-10-26 · HF API: 2026-10-26 07:15 UTC · Providers: 2026-10-26 07:26 UTC · Articles: 2026-10-26 07:27 UTC · Summary: 2026-10-07 15:01 UTC · Last verified 2026-10-26 07:27 UTC

Follow Unlimited-OCR

Watch this model

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