ThakiOCR Models

This page collects the four OCR models ThakiCloud has published on Hugging Face. All four are research releases, not product launches. They differ in input unit, required hardware, and evaluation setup, so the numbers below are not a ranking between models.

In plain terms

The four models are different tools in the same toolbox. A small screwdriver fits in a pocket and can be used thousands of times a day; a power drill is heavy but goes through a hard wall. The useful question is not which tool is better, but which one fits the job in front of you.

At a glance

This table covers size and use only. Performance numbers are listed per model below, each with its evaluation contract.

Model Size Input unit Runs on Release status License
ThakiOCR-KOEN-V1-Dense 2.13M parameters one text line (Korean, English) CPU research release Apache-2.0
ThakiOCR-KOEN-G2M-Preview 8.93M parameters one text line (Korean, English) CPU research preview, A1_GEN2_NO_GO Apache-2.0
ThakiOCR-KO-v1 0.9B-class base model + 6.39M LoRA parameters one Korean word image GPU research preview Apache-2.0
ThakiOCR-Large-27B-R2B 27B base model + 233.5M LoRA parameters one Korean word image GPU (about 53GB in bf16) research preview Apache-2.0

KO-v1 and Large-27B-R2B are released as LoRA adapters only. The base models (ATH-MaaS/OvisOCR3, Qwen/Qwen3.6-27B) must be downloaded separately at the pinned revision given on each card.

Model cards

The evaluation sets and metrics on these four cards differ, so the numbers are not a ranking between models. Each number is listed together with the conditions it was measured under.

ThakiOCR-KOEN-V1-Dense

  • Size: 2,130,249 parameters. About 1.83M of them sit in the final layer that outputs 14,152 characters.
  • Use: a small CPU baseline that reads one cropped text line. Place it after a line detector or layout model.
  • Headline metric: CER 37.14% on a real Korean development set, 17.81% if spaces are ignored.
  • Evaluation contract: 5,000 word crops from part of the validation data of AI-Hub’s OCR 데이터(공공) (Public OCR data, 71299), used for evaluation only. It is a development measurement, not an independent test set, scored by exact string match after greedy CTC decoding and NFC normalization. On held-out synthetic Korean lines from the same source as training, CER is 16.70%.
  • Speed: median 5.61ms for one 512px-wide line on 4 threads of a data-center server CPU.
  • Limitation: on 38% of real Korean word crops it inserts spaces between syllables that are not there.
  • Training data: NVIDIA OCR-Synthetic-Multilingual-v1 (CC BY 4.0) only. No AI-Hub data was used for training.
  • License: Apache-2.0 for both code and weights.
  • Links: Hugging Face · related post ThakiOCR A release
  • Release status: research release. Single-seed results.

ThakiOCR-KOEN-G2M-Preview

  • Size: 8,926,251 parameters. The same family of line-level CTC recognizer as V1-Dense, scaled up and trained again from scratch.
  • Use: research on line recognition where symbols must survive, such as won signs, dates, circled numbers, and masked digits. Median CPU latency was 13.3ms per line.
  • Headline metrics: CER 0.89% on text lines of a real held-out set, and 94.2% critical-token strict exact match on a self-generated symbol test set.
  • Evaluation contract: the real held-out set is 70 lines from one volume of the Training split of AI-Hub’s OCR 데이터(공공) (Public OCR data, 71299) that was kept out of training. The symbol test set is 5,991 newly generated synthetic lines with no text overlap with training. Both sets were opened once, after training, and the pre-set criteria were not changed.
  • Limitation: of five pre-set pass criteria (preregistered gates), it failed the date criterion (I4). It got 3 of 4 real-document date tokens right.
  • License: Apache-2.0. Training data sources are NVIDIA OCR-Synthetic-Multilingual-v1 (CC BY 4.0), in-house synthetic symbol lines rendered with SIL OFL 1.1 fonts, and the Training split of AI-Hub’s OCR 데이터(공공) (Public OCR data, 71299).
  • Links: Hugging Face · related post G2M release
  • Release status: research preview, verdict A1_GEN2_NO_GO. Do not use it to process critical fields such as amounts, account numbers, or dates in financial or contract documents automatically.

ThakiOCR-KO-v1

  • Size: a LoRA adapter for the 0.9B-class document OCR model ATH-MaaS/OvisOCR3 (rank 16, decoder only, 6.39M trainable parameters, 25.6MB).
  • Use: transcribes one Korean word image. It learned handwriting, printed text, and outdoor photographed text, and was selected so that it does not break the base model’s table and reading-order abilities.
  • Headline metric: on a fixed development set of 2,400 word crops, CER fell from 25.28% (base model) to 7.78%.
  • Evaluation contract: the development set takes 600 crops per source from the official Validation splits of the three datasets used for training. It is the same source, so it is not an independent test set. Scored by CER after greedy decoding, NFC normalization, and whitespace stripping. Document structure was checked only as “no worse than the base model”: exact table-structure match was 1.00 (base 0.98) on 50 synthetic table pages kept sealed until results were in.
  • Speed and memory: median 0.210s and peak memory 7.02GiB on one B200 at batch 1.
  • Limitation: trained and evaluated at word level only. Full-page reading quality on real documents has not been independently verified.
  • Training data: AI-Hub’s 대용량 손글씨 OCR 데이터 (large-scale handwriting OCR data, 605), 다양한 형태의 한글 문자 이미지 인식 데이터 (Korean character image recognition data in various forms, 91), and 야외 실제 촬영 한글 이미지 데이터 (Korean text images photographed outdoors, 105). The data and anything derived from it are not redistributed.
  • License: Apache-2.0. The base model is governed by its own license.
  • Links: Hugging Face · related post KO-v1 release
  • Release status: research preview.

ThakiOCR-Large-27B-R2B

  • Size: a LoRA adapter for Qwen/Qwen3.6-27B (rank 32, 496 decoder linear layers, 233.5M trainable parameters).
  • Use: research on large-model OCR of one Korean word image. Training drew words with rare characters more often.
  • Headline metric: on a fixed development set of 2,400 word crops, CER fell from 18.04% (base model) to 6.01%. Whole-word exact match is 85.2%.
  • Evaluation contract: the development set takes 600 crops per source from the official Validation splits of the three datasets used for training and is not an independent test set. Overlap with training was ruled out by file, pixel, and source-image hashes, and scoring used greedy decoding. Confidence intervals come from a stratified paired bootstrap fixed before results were seen.
  • Hardware: about 53GB of GPU memory in bf16 at batch 1, median about 0.46s per crop on one B200.
  • Limitation: the extra 0.41pp gain over the uniform-sampling adapter (R2-A) is statistically borderline.
  • Training data: the same AI-Hub datasets 605, 91, and 105 as KO-v1. The data and anything derived from it are not redistributed.
  • License: Apache-2.0. The base model is governed by its own license.
  • Links: Hugging Face · related post Large-27B-R2B release
  • Release status: research preview. No claim of superiority over specialist OCR models.

The development sets of KO-v1 and Large-27B-R2B are built in a similar way, but we have not verified that evaluation images, preprocessing, decoding, and ground-truth handling are the same. So we do not compare the two models with these numbers.

None of the four cards links a GitHub repository. Inference code ships inside each Hugging Face repository (inference.py for V1-Dense and G2M, inference_example.py for KO-v1 and Large-27B-R2B).

Which one should I pick?

Pick by input unit and hardware, not by a numeric ranking.

  • No GPU, and you need fast line-level reading: look at V1-Dense or G2M. Both take one line cropped by a line detector.
  • Lines where won signs, circled numbers, or masked digits matter: look at G2M. It is weak on dates with inner spaces and on telling ₩ from ￦, so keep a human review step.
  • You want the lightest baseline or a from-scratch architecture: look at V1-Dense. It shows what breaks on real documents when a model learns from synthetic data alone.
  • Korean words mixed with handwriting or photographed text, on one GPU (about 7GiB): look at KO-v1. It is the adapter that chose to keep the base model’s table-structure ability.
  • 53GB or more of GPU memory, and you are studying how far a large model can learn word OCR: look at Large-27B-R2B.

None of these models should read critical fields in financial, legal, or identity documents without human review. How to route documents to different models is covered in the Metis tiered document inference post, and how to commit results to business systems only after verification is covered in the Paxis verification workflow post.

For research use

  1. Download from Hugging Face. Each card’s Files table lists weights, configuration, and inference code. For adapters, pin the base model revision given on the card.
  2. Reproduce with the card’s inference code first. Nothing outside the card’s preprocessing, prompt, and decoding settings was evaluated. V1-Dense and G2M can be checked right away with python inference.py --demo.
  3. Read the documentation. Failure types and evaluation set composition are in each card’s Known limitations section and in the related posts. Why the four models come in different sizes is explained in Four Open Korean OCR Models Compared.

For adoption review

  1. Define the use case. Write down document types, input unit (line, word, page), daily volume, and the fields that must not be wrong. None of the four models finds text positions on a page, so a separate detection step is needed.
  2. Check security and on-premises conditions. The weights can be downloaded and run inside a private network. Running training and serving inside the customer’s environment is covered in the Maxis training post and the Metis tiered inference post.
  3. Send a technical inquiry. For evaluating on your own documents or deploying on the platform, contact ThakiCloud.

In development

These models are in development. No performance numbers are published. ThakiOCR-KO-v2: there was not enough document and handwriting data for an independent evaluation, so the official evaluation was not run and the release is on hold. It is not on hold because of poor performance. ThakiSight-Edge: training is in progress.

The list below shows every post tagged thakiocr. The main ones are:

48.1% and 4x: The Day Models Started Fixing Models

MIT CSAIL used GPT-6 Astra to edit Qwen’s weights directly and pushed its math score to 48.1%, and the same day Gemini 4 Carbon posted a 90.4% mean at four times lower cost than Opus 5.5.

Same Day, Two Agents, Neither Moved

DeepSeek’s ‘Super Agent’ and BespinGlobal’s ‘HelpNow Computer’ launched on the same day, and both chose ‘observe first’ as the default.

Hi3D 1.0: a production-validated image-to-3D model, now open source

The model formerly known as Sparc3D went open source as Hi3D V1.0: a self-pruning sparse VAE that refines coarse geometry to 1536-resolution with reference images, trained on about 2 million curated 3D assets, with a technical-report claim that 90 percent of outputs pass 3D-print slicing.

The Day Markdown Became an Executable File

A web page only has to drop a single AGENTS.md into the Downloads folder, and the coding agent working in that folder starts following the attacker’s instructions.

The Autonomy Ladder: When Kubernetes incidents are stratified by remediation complexity from L1 to L3, the boundary where unattended remediation is safe and cheap reaches L2 in the nominal scenario, and that boundary and the per-incident cost can be computed before any permission is granted

How far can an LLM agent carry a Kubernetes incident without human intervention, what does it cost per incident, and from where must a human stay in the loop? ThakiCloud’s analytic paper The Autonomy Ladder stratifies incidents by remediation complexity from L1 to L3 and scores diagnosis, planning, execution, and verification as a closed loop.

Cortex: Consolidating Context into Parametric Memory

Cortex anticipates the questions a context will be asked, distills only those behaviors into the FFN, and after 50 sequential model edits recovers 88.9% of the context gain while retaining 99.8% of the original performance.

The Inference Moat Was Filled

The open frameworks SGLang and vLLM have caught up with DeepSeek’s newly open-sourced inference engine DeepInfer, matching or exceeding its performance.

The Morning the Black Box Myth Ended

Research on trillion-parameter models has confirmed that the tangled knots inside neural networks are far simpler than previously thought.

The Model Is No Longer the Menu

From the mandatory 50/30 domestic-model mix in Korea’s ‘Everyone’s AI’ to KakaoTalk’s model routing, today’s 18 news items point at one place.

Number One at 19.2%

Zai’s GLM-Legal took the number one spot in open-weight legal AI, but the score that number one earned was 19.2%.

The Day Problems Became Property

OpenAI’s 700-paper release, integer multiplication improved by 69.13% in a single day, and a mathematician’s warning to stop publishing.

The Day the Pocket Becomes the Data Center

The blueprint for a 100-billion-parameter phone by 2028, the all-night arithmetic of a Mac running 3.4x faster, and the $30-billion building that capital markets turned their backs on.

Autonomy: Show the Receipt

In October, SAP officially released its ‘Autonomous Enterprise’ and switched billing to a consumption model of ‘AI units.’ In the same week, Samsung SDS outlined its four elements of agent governance, SK AX unveiled a control tower, and KT said ‘use it often.’ The model layer is flooding with $12 billion and a trillion parameters.

Harness Engineering: A Source-Code Study That Autopsied 11 Coding Agents

Do you need a framework to build a coding agent? Wavestone AI Lab read roughly 4 million lines across 11 production agents and found the opposite: none of them uses one, and the 83-page paper details independent design over 7 canonical subsystems plus pattern diffusion with a weekly half-life.

The Morning Agents Wore Two Badges

About 25,000 Shinhan Bank customers’ information leaked, and the intrusion channel was not the core app but a loan-solicitor-only service.

Two paychecks: OpenAI at $500, Meta at $0

OpenAI put a $500-a-month tier on its 24-hour autonomous agent ‘Dots,’ and Meta expanded ‘Muse’ small-business integrations, mostly for free.

How One Canceled Launch Moved the Terrain

The week GPT-6.1 Astra’s October launch was pulled over safety concerns, the layers actually doing the work had already moved down to mid-tier models, agents, and subscription contracts.

The AI That Took the Night Shift

The National Intelligence Service’s unmanned AI cyber defense team, Project Phalanx, has been put into live action, and Jensen Huang has declared cybersecurity the next growth axis for AI.

The Week Keys Moved Faster Than Audit

The name of the folder OpenAI’s agent created on its own is ‘LOOT.’ In the week that 24 incidents accumulated and a freeze on advanced models was ordered again, the industry kept handing the keys to email and bank accounts to agents.

Two Verdicts, the Same Word: Control

A court ruled that a model’s safety device is a national security threat, and the CEO of the same company warned that agents without safety devices threaten humanity.

The Exam Room Had No Door

Grok 4.7 bypassed the network restrictions during a coding benchmark and pulled in an existing solution from outside.

The Day Models Got Passports

In the same morning, OpenAI urged the United States to lead a global standard with independent model audits, while China is probing DeepSeek and Moonshot over 35 million leaked data requests.

Audit Passed. The Data? In the Clouds.

The day the privacy regulator made the ISMS-P audit stricter, the inspector asked one thing: ‘Where’s the data?’ ‘In the clouds, sir.’

Safety Is No Longer a Research Question

A $2 trillion IPO delay request from a pension investor, a Big Three safety body formed without an antitrust exemption, and a $6.87 million resident evaluator.

When the Recipe Became Free, the Kitchen Became the Brand

What connects the four-legged robot and the airport check-in robot is a common execution environment called ‘Airpath.’ From Moonshot AI’s field engineers, to KT’s token factory, to the hyperscaler’s power plant.

Giving Outsiders Employee-Level Access

OpenAI pledged external safety reviewers the same access as its employees, and the week Anthropic granted METR permanent employee-level access, it blocked a scientist’s account over bioweapon risk.

Search Is Translation, Not Matching

The “the data is there but the user cannot find it” failure is almost never a storage failure; it is a translation failure.

Security: Not the Next Market, but the One Already Clocked In

Cybersecurity is the market Jensen Huang called the ‘next great AI market.’ On the same day, Naver put up a 200 billion won bet on a security-specialized model, and Naver Cloud named cybersecurity the next stage of sovereign AI.

What the one-line prompt removes, and what the code measures

Does one line de-AI your text? We put Anthropic’s officially documented mannered-prose one-liner through a 3-round experiment: metaphors and flourishes fell from 28 to 0, and the comma-rhythm tells that survived were the code gate’s job.

Rogue Agents Have a Case Number Now

OpenAI has filed a formal incident report with the European Commission over autonomous agents that took over a German website.

It Saw the Oar Before It Fell

A fast, experimental vision model spotted a whale, a UFO, and a mustachioed fish in a perfectly calm lake.

The Handshake Became Its Own Conference

The AI managed-service-provider wars are really a conference arms race, and the answer to ‘whose server room?’ is an invitation to the next conference.

The Day AI Moved in Two Directions at Once

On the same day, OpenAI shipped a 10-trillion-parameter model and its own chip, Anthropic pitched a $30 trillion market, and Perplexity put an entire agent runtime into one box on a desk.

The Day We Made Agent Work a 'Protocol'

We installed the new cookbook that wires Anthropic’s Claude Managed Agents to the AG-UI protocol and CopilotKit, and ran a hands-on experiment with HITL interrupt-resume flows at the protocol level.

Your Time Bug Is Actually a Type Error

A settlement off by a day, a batch job that ran twice, timeouts firing in a storm for no visible reason: these look like unrelated incidents.

We Paid Extra to Look Worse

Thirty seconds of 480p cost $4.12, and a good chunk of that went into making it look grainier.

Hidden Reasoning Was Not Hidden

In the same week researchers recovered 62 API keys from 315,320 reasoning blocks, a tool that strips provenance watermarks widened its supported formats.

The Day One Company Braked, Everyone Else Accelerated

OpenAI announced it would slow the pace of frontier model development, and in the same 24 hours Alphabet raised 25 billion dollars and MiniMax open sourced the weights of a 33 billion parameter model.

Building an AI-Native CI/CD Pipeline as a Solo Developer

For solo developers and small teams with no colleague to review their code, this post builds a pipeline that hands testing, review, and deploy decisions to AI agents, covering everything from gate design to rollback alerts with real configuration.

Thank You in Seven Languages

This short video teaches thank you in Korean, English, Japanese, Chinese, French, Russian and Arabic.

Sourcing Every Claim

You can read tens of thousands of chemistry papers, and still ask one thing.

An Evaluation That Trusts No Single Number: Building a Sovereign LLM-Judge Service on Binary Decomposition and Deterministic Gates

Using an LLM as a grader, the practice known as LLM-as-a-judge, is now the default in model development, but the evidence that piled up through 2026 shows that a scalar judge producing a single score is fragile to prompt wording and answer position, drifts toward the middle of the scale, and collapses to coin-flip reliability against adversarial inputs.

Editing Video With a Coding Agent: A Look Inside the video-use Skill

Shared by midudev and quickly making the rounds, browser-use’s video-use is a free, open-source skill: drop raw footage into a folder, type one sentence, and a coding agent handles cutting, filler removal, subtitles, color grading, animation, and rendering.