新智元报道
几周前，DeepSeek用V4.1-Flash再次告诉全世界：顶级模型，不一定要天价！
今天，轮到OCR了。
面壁智能开源的MiniCPM-OCR 1.0，只有1.3B参数，却同时拿下解析能力和效率两项第一，把一众大厂模型甩在身后。
这是MiniCPM V/o 家族首个OCR专用模型。
但这，还远不是这次发布最狠的地方。
真正打破旧秩序的，是同一天开源的另一个模型——MiniCPM-OCR Jev。
奇怪的是，作为OCR新王身边的「红人」，它竟然连一个字都不吐。不答题，只判卷。一个OCR新王，为什么要给自己配一个专职的「裁判」？
这个反常的举动，恰恰戳破了整个AI圈心照不宣的秘密。
如腾讯HunyuanOCR-1.5、阿里系/智谱系大模型等一众大厂模型，纷纷发力文档解析OCR领域。
在这个大模型动辄将参数堆到几百B（十亿）、算力如流水的年代，面壁智能开源了MiniCPM-OCR 1.0。它只有 1.3B 的参数，完成了一次教科书般的降维打击。
MiniCPM-OCR 1.0杀出重重包围，几乎撑爆了整个多边形的边界。
在6项核心测试中，它以碾压之势拿下了4项SOTA（业内最佳）：
- Real5-OmniDocBench：93.68（全场最高）。这可不是实验室里的高清扫描件，而是专考手机随手拍、镜头畸变、反光翻拍的「魔鬼考场」。
- Wild-OmniDocBench：92.39（全场领跑）。专治五花八门的野生真实文档，远超同量级的竞品。
- OCRBench v2：54.7（绝对压制）。在通用字符识别与关键信息抽取上，把第二名（47.0）远远甩在身后。
- CHAOS-Bench：29.78（断崖领先）。这项测试考的是忠实度，在强语义先验噪声的干扰下，模型能否忠实于图片进行识别，最终它的得分是第二名的1.4倍以上。
MiniCPM-OCR 1.0晋升为开源OCR新王。
MiniCPM-OCR 1.0文档解析能力一骑绝尘，支持中英双语文档与 PDF 解析，兼顾标准文档和复杂采集条件，在扫描、弯曲、拍照、倾斜及翻拍等场景下保持稳定表现。
话不多说，咱们先来上实测，看看这个MiniCPM-OCR究竟有多能打？
比如国外网友手写的剧本，几乎秒杀：
再比如，多字体的英文识别，也轻松拿下。
再比如，我们生活中常见的发票，到底有多难认？
红色表格线、两枚压字的印章、密密麻麻的数字和乱码，几乎集齐了文档识别的所有「坑」。
MiniCPM-OCR 1.0的答卷是：结构一层没塌。
购买方、销售方、密码区左右分栏，合并单元格原样还原；发票代码、校验码、18 位纳税人识别号、16 位银行账号，这些错一位就作废的长串数字，逐位对上。
连压在字上的红章，它也知道单独切出来，不让印泥糊进正文。一张图丢进去，吐出来的就是一份结构化的电子表格。
再让MiniCPM-OCR 1.0识别下学霸笔记。拍的时候一道斜阳正好打在纸上，半页泛着橙光。
内容是用费马的「虚拟相等法」求抛物线最大值：一边推公式，一边画曲线，箭头、批注、方框挤成一团。
MiniCPM-OCR 1.0 把整条推导链逐行还原成了标准 LaTeX。带分数线的约分、上下标，一步不乱，最后落在圈出来的结论 b − (x₁ + x₂) = 0。
还有两张手绘示意图被单独切成图片保留。即便是光线再乱、字迹再潦草，推导的逻辑一步没丢。
下面这张被揉皱的英文教材复印页，出自Gilbert Strang的经典教材，满页都是矩阵和不等式。
MiniCPM-OCR 1.0先把图片自动摆正2°，再逐行还原。分块矩阵 [A −I]、上标 x*、下标 x₁，全部转成可直接渲染的LaTeX。
还有第7题的4×4矩阵，一个数都没错。一张皱巴巴的复印纸，就这样变回了一页干净的电子教材。
下面加大难度，让它去识别一个随手拍下的一页杂志。
一张涂鸦墙满版照片、几种字体的大标题挤在一起。最难的一块，是印在砖墙涂鸦照片上的白色小字。
背景花得几乎看不清，MiniCPM-OCR 1.0 还是把「Getting real」这段的四个自然段完整读了出来，照片则被单独切成图片。
一页杂志拍下来，就是一篇能直接复制、翻译、检索的文章。
再来看下面这一整版中文的日报，八篇稿件挤在同一张版面上：横排、竖排标题混搭，正文切成密密麻麻的窄栏，新闻照片穿插其间，还有「下转第四版」的跨版提示。
MiniCPM-OCR 1.0 像一位老编辑，把这张报纸重新拆了一遍版。
连报头下方的刊号、邮发代号、新闻热线和电子邮箱，这些小字也一个没漏。
一番实测下来，MiniCPM-OCR 1.0的表现确实够硬核。
如果说精度是入场券，那效率就是企业买单的最终理由。
以每秒处理页数计算：
MiniCPM-OCR 1.0以1.217页/s的吞吐量登顶（1.00x）。
对比之下，老牌劲旅PaddleOCR-VL-1.6只有0.818（0.67x）。
而某些主打高精度的模型，如MinerU2.5-Pro-Advanced仅有 0.336（0.28x），TeleOCR更是惨跌至0.141（0.12x）。
快，不是靠堆硬件堆出来的。
MiniCPM-OCR 1.0 用 LLaVA-UHD v4 的视觉分片方案，把超大图像切块高效编码；再叠加 DSpark 投机采样解码加速，端到端处理效率直接提升 55%。
小参数、轻架构、快解码，三板斧叠在一起，才有了单张RTX 5090上平均不到1秒解析一页的速度。
在账本上，吞吐量就是真金白银。在这个指标上，MiniCPM直接拉低斩杀线。
这不仅是成本上的降维打击，更是让顶尖AI能力真正走向「技术普惠」。
而这，恰恰是面壁智能一直在做的事。
从把GPT-4V级多模态能力塞进手机的MiniCPM-V，到今天1.3B的MiniCPM-OCR，面壁始终在用更小的模型，干更大的活。
MiniCPM-OCR，不过是这条路线上又一次「以小博大」。
过去，动辄需要集群算力才能支撑的工业级文档解析管线，现在被压缩进了一张消费级显卡甚至边缘端侧设备中。
中小微企业、科研团队和独立开发者无需承担高昂的算力税，在本地就能搭建起私密、安全的高吞吐生产流，让顶尖的 AI 能力真正变成人人可享的「平民工具」。
但最狠的还不是准确率和效率。
OCR的下一战，不再是识别得更准，而是识别得更可信。
过去几年，大家都在宣称自己的大模型能做到 99%。
但问题在于，视觉语言模型天生带有「幻觉」。
在金融核保、法务合同、医疗病历这些极为严苛的业务里，哪怕只有 1% 的不可控风险，都可能招来千万级的合规灾难。
问题的本质是信任。
假设100张发票里，AI识别错了1张，但你不知道是哪一张。你会怎么办？
答案很绝望：你只能雇人，把这100张发票从头到尾、一张不落地重看一遍。
于是，人类商业史上最荒诞的一幕出现了：企业花重金购买了最前沿的AI算力，可能转头又雇了几十号人，像流水线工人一样，把那99%正确的数据人肉核对一遍。
不知道错在哪的99分，和0分一样，最终都需要人类来兜底。
拉长视野，你会发现过去一年，文档OCR模型的参数规模正在经历一条陡峭的「缩水曲线」。
从3B到1.2B，再到0.9B、0.6B，开源越来越密，技术门槛被一路拉平。
技术架构的进步提升了「识别能力」，「认字」这件事，正在从技术稀缺品变成算力日用品。
当人人都能认字时，真正稀缺的变成了什么？是可信。
这一次，面壁智能第一次把「可信」从准确率里单独拎了出来，立成一个全新的品类：可信 OCR。
在此之前，行业里根本没有这个概念。所有厂商都在比谁「认得准」，却没有人去回答一个更要命的问题：这一份结果，到底能不能信？
而就算有人想解决「可信」，过去行业的思路也非常单线：大力出奇迹。把排版、公式、图表解析，连同「检查自己有没有认错」的任务，全部塞进同一个巨无霸视觉语言模型里。
但这违背了第一性原理。
想想你自己写过的稿子。自己检查十遍，错别字依然会溜过去；但同事扫一眼，可能立刻就指出来。原因不是你不认真，而是你的大脑会自动「脑补」你想写的东西。
新一代OCR模型也是一样。无论参数做多大，在面对自己脑补出来的幻觉时，它天然缺乏概率校准能力。让大模型给自己打分，就像让考生给自己判卷，它永远盲目自信。
市面上一些OCR引擎附带的「置信度」，本质上也是这种自己给自己打的分。
人类社会几千年前就想明白了这件事：会计记账，必须有审计查账；运动员比赛，必须有裁判吹哨。
面壁智能的破局之道，就是分工协作：解析模型只管干活，决策模型专职验货。
MiniCPM-OCR 1.0专注识别，决策模型MiniCPM-OCR Jev负责「识别结果的可信度」。
MiniCPM-OCR Jev对文本、公式和表格候选评分
「感知生成」与「理性裁决」在AI架构中第一次被彻底拆开。
这也是业内第一次，有人把解析模型和决策模型打包一起开源。别人发布的是一个OCR模型，面壁发布的，是一条自带质检的OCR流水线。
能力可以集中，信任必须分权。
MiniCPM-OCR Jev自带三个极其硬核的底层设计。
第一，告别盲目自信，只输出「统计学真相」。
MiniCPM-OCR Jev给出的置信分数，经过大量真实人工标注数据进行统计校准。
大模型的「幻觉」，被量化成了一个可以核对的数字。
第二，它快到几乎不占时间。
过去让大模型做质检，它得字斟句酌地生成一大段推理，延迟极高。
而MiniCPM-OCR Jev拿到OCR候选结果，只做一次前向计算，不生成 任何文字，毫秒级给出裁决。
它就像高速公路上的 ETC 闸口， 卡在「文档识别和业务入库」之间，让90%以上毫无疑问的数据「秒过」 。
第三，小身板，拥有巨头级的判断力。
在衡量判别能力的 AUROC 指标上，仅2B参数的MiniCPM-OCR Jev拿下了93.6。
作为对比，参数高达397B的通用大模型Qwen3.5是93.8。
它以极微小的算力代价，几乎打平了超大模型的审查能力。
若是几种办法都试完还不过线，就标出来交给人复核。
拿来试的，是DeepSeek-V3技术报告里那张训练成本表。557.6万美元的出处，就是它。
第一遍，数字一个没错：2664K、119K、5K，连合计的5.576M都在。但四个美元符号全被读成了公式标记，合计金额变成了「)5.576M」。
光扫一眼数字，人很容易放过去。Jev没放过，只给了约0.000001分。
程序当即把表打回重读。第二遍，原图放大两倍、四周留白，美元符号还是不对，分数纹丝不动。
第三遍，程序按预设策略换了一套识别流程。这回四个「$」全部回来了，Jev复判给到0.9692，重试停止，表格存成JSON。
我们把最终结果逐项对回原图：表头、各阶段GPU小时数、美元金额，全部一致。
三轮识别、两次返工，中间没人手工补一个符号，也没人守着点「再试一次」。OCR负责读，Jev负责判，程序决定要不要重来。
落到真实的商业运转中，这是一场摧枯拉朽的效率革命。
系统把高置信数据直接放行，只把剩下的可疑切片推给人工。
诺贝尔经济学奖得主丹尼尔·卡尼曼在《思考，快与慢》中提出，人类有两套系统：System-1 是快思考（直觉），System-2 是慢思考（理性推理）。卡尼曼认为，直觉常常出错，需要理性去纠正它。
但 MiniCPM-OCR Jev极其精妙地把这个故事反转了。
它是毫秒级的System-1，去给费时费力的生成模型System-2把关。
为什么它的「直觉」那么准？因为它是由海量真实数据一遍遍喂出来的。就像一位干了三十年的老银行柜员，钞票在手里一搓就知道真假。他不需要长篇大论，他的「手感」本身，就是最严密的统计规律。
需要特别指出的是，Jev模型并不只适配MiniCPM-OCR。
作为一个高度解耦的独立决策模型架构，Jev可以作为即插即用的「置信度外挂质检员」，灵活架设在任何一个OCR之后。
无论你的业务流水线目前使用的是经典的 PaddleOCR、垂直领域的 TeleOCR，还是调用大厂闭源的商业 API，甚至是你本地私有化微调的文档解析模型，无需改动原有的识别底座，都可以将 Jev 嵌入为一道校验网关。
它以统一的无偏标准对各路异构输出进行校准判别，成为整个企业级文档处理管线中通用的「可信过滤器」。
AI 的上半场，比谁更聪明；下半场，比谁更诚实。
信任不是「永远不出错」，而是「错了，我知道」。
这次面壁智能的开源，把行业门槛降到了令人发指的地步：两个模型权重全部开源，首日即支持Hugging Face、vLLM、Ollama，甚至配有不用配环境的在线WebDemo。
站在更大的产业视角来看，MiniCPM的OCR与Jev组合，正在掀起一场更深远的工程革命：
OCR模型的定位，已经全面升级为AI Agent「上下文治理」的核心基础设施。
当下的AI Agent正在接管越来越多的专业工作流，而智能体思考的质量，直接受限于输入上下文的纯度与结构。
MiniCPM-OCR解析出的高质量排版、富文本与Markdown，直接作为Agent最核心的上下文输入。它让智能体无需在充满乱码和格式破碎的泥潭中盲人摸象，为端到端自动化扫清了第一道认知障碍。
在各行业核心业务中，大量「非结构化文档转存数据库」是刚需。
比如，财务审计与核算领域成千上万页资产负债表、利润表、发票底稿，需毫秒级结构化提取并对齐条目。
原有的通行做法非常繁复：
先用 OCR 把图片转成文本或 Markdown，再调用大语言模型（LLM），配上长 Prompt 和特定 JSON Schema 规则做「二次蒸馏提取」，最后才存入数据库。
这带来多一倍显存占用、高额API账单和数秒延迟，更埋下二次幻觉隐患。
MiniCPM-OCR 的原生关键信息抽取KIE改变了这一点——端到端一步到位。
无需在后端再挂一个庞大LLM，即可精准识别关键实体与字段关系，把文档流直接对接企业级存储引擎。
放到更大的产业图景里看，这条流水线来得正是时候。
全球OCR市场仍在高速增长。据The Business Research Company统计，市场规模将从2025年的166.6亿美元增长到2026年的192亿美元，2030年有望达到339.2亿美元。
但增长的逻辑已经变了。
过去，企业买单是因为「更多文档需要被数字化」；现在，企业要的是数字化之后，结果「能被信任、能被自动化」。金融保险与医疗健康这两个对错误零容忍的行业，合计贡献了超过四成的市场份额。
与此同时，「透明」和「可审计」正在从加分项变成准入门槛。一个只会输出结果的解析模型，回答不了「这份结果能不能直接用」；一个经过校准、敢说「我不确定」的决策层，才是下一阶段的入场券。
当所有人还在卷准确率时，面壁已经提前站到了下一个路口。更重要的是，这一次，「可信」也被开源了。
人类文明最了不起的发明之一，不是让某个人进化得完美无缺，而是建立了一套人与人之间互相监督、彼此校准的制度。
现在，轮到AI走向文明了。
让校验隐身，让信任显形。
编辑：大卫 桃子 摩西
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
A few weeks ago, DeepSeek told the world again with V4.1-Flash: Top model, not heavenly price!
Today, it's OCR's turn.
MINICPM-OCR 1.0 with only 1.3B parameters for the face wall ' s smart open source, while taking the first of the two, the ability to solve and efficiency, tossing the large plant model behind its back.
This is the first OCR-specific model of the MiniCPM V/o family.
But that's far from the worst part of publish this time.
What really breaks the old order is another model of the same day's open source, MiniCPM-OCR Jev.
Strangely, as the "red man" next to the new king of OCR, he didn't even spit a word. No questions, just papers. Why would a new king of the OCR give himself a full-time referee?
This unusual move is precisely the breaking of the secret of the whole AI circle.
TencenthunyuanOCR-1.5, Arith/Zhipu are models of large plants, such as LLM, which are active in the analysis of the OCR field.
In the age of LLM, which tends to pile parameters to hundreds of B (billions of dollars) and computes to flow water, the face wall is open to MiniCPM-OCR 1.0. It had only 1.3B parameters and had completed a textbook-like downfall.
MiniCPM-OCR 1.0 has killed many encirclements and almost blew up the entire polygonal border.
Of the six core tests, it took four SOTAs by crushing:
- Real5-OmniDocBench: 93.68 (highest in the field). This is not a high-resolution scan from the lab, but a "demon exam" dedicated to hand-to-hand video, shot malformation and reflection.
- Wild-OmniDocBench: 92.39 (whole lead). Specializing in a wide variety of real wild documents, well above the level of competition.
- OCRBench v2:54.7 (absolute repression). On generic character recognition and the extraction of key information, the second name (47.0) is left far behind.
- CHAOS-Bench: 29.78 (crawling ahead). The test is faithfulness, and the model, with the interference of a priori noise in the semantics, is faithful to the picture for identification, and ultimately scored more than 1.4 times the second.
MiniCPM-OCR 1.0 was promoted to the new opening-source OCR.
MINICPM-OCR 1.0 document resolution capability is one-sided, supporting the analysis of the Chinese-British bilingual document with PDF, balancing standard documents and complex collection conditions, and maintaining stable performance in the scanning, bending, photographing, tilting and flipping scenes.
Let's get to the bottom of this, see how good it is to fight MinicPM-OCR.
For example, a script written by a foreign online friend, almost a second:
And, for example, multiple fonts for recognition in English are easy to take.
And for example, how hard are the invoices that are common in our lives?
The red table line, the two typed seals, the coded figures and the hyphenation, have almost all the "holes" identified in the document.
The answer to MiniCPM-OCR 1.0 is that the structural layer is intact.
Purchaser, seller, password area, right-and-right column to consolidate cells; invoice code, calibration code, 18 taxpayer identification number, 16 bank account numbers, each of which is invalid with a long string of numbers, one by one.
Even the red stamp on the word, it also knows to cut it out alone so that it doesn't leave mud in the text. A map was thrown in, and it was a structured spreadsheet.
In addition, MiniCPM-OCR 1.0 will be able to identify the writings of the school. It's like a sun on paper, half a page of orange light.
The content is to use Ferma's "virtual equivalent" to get the maximum value of the parabolic line: push the formula while drawing curves and crowding out arrows, notes and boxes.
MiniCPM-OCR 1.0 restores the entire chain to standard LaTeX. The approximate, upper and lower mark with a fraction line, one step at a time, ends up in circles, b − (x1 + x2) = 0.
Two hand-drawing diagrams were kept in separate photographs. Even if the light goes up and up and down, the logic behind it is not lost.
Here's a wrinkled copy of the English textbook from Gilbert Strang, full of matrices and variants.
MiniCPM-OCR 1.0 automatically corrects the picture at 2° and returns the picture line by line. Block matrix [A - I], superscript x*, subscript x1, all converted to direct rendering of LaTeX.
And there's the 4x4 matrix for question 7, which is the right number. A wrinkled copy of paper was turned back into a clean e-book.
It makes it harder down there to identify a side-by-side magazine.
A graffiti wall full of photographs and several large titles in fonts. The hardest piece is the white little word printed on the picture of the brick wall.
The background is almost blind, and MiniCPM-OCR 1.0 is still reading the four natural parts of the "Getting real" section, and the photographs are cut into separate pictures.
One page of the magazine was filmed as an article that can be reproduced, translated and retrieved directly.
In addition to this whole Chinese daily newspaper, eight articles are on the same page: a mix of rows, vertical headings, a narrow text cut into a tight wall, a news photo in between, and a cross-version reminder of the "Rounddown 4" page.
MiniCPM-OCR 1.0, like an old editor, reopened the paper.
Even the issue numbers, post code numbers, news hotlines and e-mails at the bottom of the newspaper header remain intact.
Indeed, MiniCPM-OCR 1.0's performance is strong enough.
If precision is a ticket, efficiency is the ultimate reason for the company to pay.
Based on the number of pages per second:
MiniCPM-OCR 1.0 peaks at 1.217 pages/s (1.00x).
In contrast, PaddleOCR-VL-1.6 of the old-style Power Brigade is only 0.818 (0.67x).
Some models with high precision, such as MinerU2.5-Pro-Advanced, have only 0.336 (0.28x), and TeleOCR has fallen to 0.141 (0.12x).
Come on, it's not from piles of hardware.
MiniCPM-OCR 1.0 uses the LLAVA-UHD v4 visual fraction scheme to encode super-heavy image slices efficiently; super-heavy with the DSpak speculative sample decoded faster, increasing direct end-to-end processing efficiency by 55%.
Small parameters, light structures, fast decode, three axes folded together, allowing for a single RTX 5090 at an average speed of less than one second to resolve a page.
In the books, throughput is real silver and silver. On this indicator, MiniCPM directly lowers the kill line.
This is not only a cost-cutting blow, but also a real "technology inclusion" of top AI capabilities.
And that is exactly what face-wall intelligence has been doing.
From MiniCPM-V, where GPT-4V multi-modular capability is inserted into a mobile phone, to MiniCPM-OCR, where 1.3B is today, the wall has been doing more work using smaller models.
MiniCPM-OCR is just another "little big" on this route.
In the past, industrial-scale document resolution lines, which often required cluster computing to support them, were now compressed into a consumer-grade GPU or even marginal equipment.
Small and medium-sized enterprises, research teams and independent developers can build private, safe high-volume production streams locally without high computing taxes, making the top AI truly a civilian tool for everyone.
But the worst is not the accuracy and efficiency.
The next battle of OCR is no longer better identified, but more credible.
Over the past few years, everyone has been claiming that they can do 99% of their LLMs.
But the problem is that the visual language model is born with an illusion.
Financial security, legal contracts, medical records, even one percent of the uncontrollable risks in extremely demanding operations, could attract tens of millions of degrees of compliance.
The essence of the problem is trust.
Assuming that out of 100 invoices, AI identified the wrong one, but you don't know which one. What would you do?
The answer is desperate: you can only hire people to look at the 100 invoices from scratch.
So, the most ridiculous scene in human business history is that firms spend heavy money on the front-line AI calculator, and they probably turn around and hire dozens of people, like water line workers, to check the right 99 percent of the data.
I don't know what's wrong at 99 points, like zero points, and ultimately humans are needed to cover it.
Longer horizons, you'll find that over the past year, the size of the parameters of the OCR model of the document is going through a steep "shrunk curve."
From 3B to 1.2B, to 0.9B, 0.6B, the open source is becoming more and more dense, and the technical threshold is evened.
Advances in the technological architecture have enhanced the "identification" and the "literacy" phenomenon is moving from technology scarcity to computing.
When everyone can read, what really scarce becomes? It is credible.
This time, for the first time, the face-wall intelligence singled out "credible" from the accuracy rate as a brand-new type: credible OCR.
Until then, there was no such concept in the industry. All the manufacturers are "recognizing" more than anyone, but no one is going to answer a more serious question: can you believe this?
And even if someone wanted to solve the "credible" problem, the past was a very simple one: a miracle. It's like putting all the layouts, formulas, graphs and graphs, together with the task of "checking if you're wrong" into the same giant visual language model.
But this runs counter to the principle of firstness.
Think about the articles you wrote yourself. Ten times you check yourself, and the wrong words still slip away; but a colleague takes a quick look and may immediately point it out. It's not because you're not serious, it's what your brain automatically does.
So is the new generation of OCR models. No matter how large the parameter is, it naturally lacks the ability to calibrate probabilities in the face of an illusion of its own brain. Let LLM score for himself, like let the examiner pass his paper, and he'll always be blind.
Some of the OCR engines on the market have a "confidence" that is essentially a self-rated rating.
Human society had come to understand the matter thousands of years ago: accounting, having to be audited; sportsmen, having to be judged.
The solution to the problem of face-wall intelligence is division of labour: the analysis model works only, and the decision model specializes in inspection.
MINICPM-OCR 1.0 focused on identification, and the decision model MINICPM-OCR Jev was responsible for "validation of results".
MiniCPM-OCR Jev candidate for text, formulae and tables min
For the first time in the AI structure, "the creation of perception" and "rational decision" were completely dismantled.
It was also the first time in the industry that the pyrolysis and decision-making models were packed together. The other publish is an OCR model, the face wall publish, a self-checked OCR flow line.
Capacity can be centralized and trust must be decentralized.
MiniCPM-OCR Jev carries three very hard core bottom designs.
First, say goodbye to blind self-confidence and only to "statistical truth".
The confidence scores given by MiniCPM-OCR Jev were statistically calibrated through a large number of authentic manual markers.
LLM's "discretion" was turned by quantization into a verifiable figure.
Second, it's almost out of time.
In the past, LLM had been allowed to perform a qualitative examination, which had generated a large amount of reasoning, with a very high delay, word for word.
And MiniCPM-OCR Jev gets the OCR candidate, does a forward calculation only once, does not generate any text, and gives the verdict in milliseconds.
It's like the ETC gate on the freeway, stuck between "document recognition and business library", giving more than 90% of the unquestioned data "seconds."
Third, small body, with a big head of judgment.
For the AUROC indicator, only 2B parameter MiniCPM-OCR Jev took 93.6.
In contrast, the generic LLMQwen 3.5 for parameters up to 397B is 93.8.
At a very small calculator cost, it almost evened the review capability of the super-LLM.
If several approaches are exhausted, they are marked for review.
It's a training fee sheet from the DeepSeek-V3 technical report. The source of $5.576 million is it.
For the first time, the figures were correct: 2664K, 119K, 5K, even the sum of 5.576M. However, all four United States dollar symbols were read as formula markers, and the total amount became ") 5.576M".
Take a look at the numbers, it's easy to put them through. Jev didn't let it go, he only gave about 0.0001 points.
The program immediately rereads the watch. For the second time, the original figure was twice as large, the surroundings were white, the United States dollar symbol was still wrong and the fraction lines remained intact.
Thirdly, a process of identification was replaced by a predefined strategy. This time, all four "$" came back, and Jev was reinstated until 0.9692, retry stopped, and the form was stored as JSON.
We turn the final results to the original chart: the table header, the number of GPU hours at each stage, the dollar amount, all the same.
Three rounds of identification, two returns to work, none of which adds a symbol by hand, and no one holds the point "tried again". OCR reads, Jev judges, and the procedure decides whether to start over.
It's an efficient revolution that destroys the real business.
The system releases high confidence data directly, pushing only the remaining suspicious slices to manual.
The Nobel laureate for economics, Daniel Kanyman, in Thinking, Speed and Slow, argued that humanity has two systems: System-1 is fast thinking (intuitive), and System-2 is slow thinking (reasonable reasoning). In Kaniman ' s view, instincts often go wrong and need reason to correct them.
But MiniCPM-OCR Jev turned the story upside down.
It is a millisecond-grade System-1, to give a time-consuming and labour-intensive model of production, System-2 off.
Why is it so "intuitive"? Because it's being fed over and over again by a lot of real data. Just like an old bank clerk who's been in business for 30 years, he knows the truth at once. He doesn't need a long speech, and his "touch" is itself the closest statistical rule.
It should be noted, in particular, that the Jev model does not fit only MiniCPM-OCR.
As an independent decision-making model structure with a high degree of decoupling, Jev can act as an “out-of-trust” checker, with flexibility to post any OCR.
Whether you use the classic PaddleOCR, the vertical area TeleOCR, or even your locally privatized fine-tuning document resolution model, you can embed Jev into a check gateway.
It calibrates all road isomer output using a uniform non-event standard and serves as a "credible filter" throughout the enterprise-level document processing line.
The first half of AI is smarter than who; the second half is more honest than who.
Trust is not "never wrong" but "wrong, I know."
The open source of this wall of intelligence has brought the industry threshold down to an appalling level: Both models have full open source weights, with the first day supporting Hugging Face, v LLM, Ollama, and even an online WebDemo with no environment.
From a larger industrial point of view, MiniCPM’s OCR and Jev combination are creating a far-reaching engineering revolution:
The positioning of the OCR model has been fully upgraded to the core infrastructure of AI Agent “context Governance”.
The current AI Agent is taking over an increasing number of professional workflows, while the quality of AI agent thinking is directly limited to the purity and structure of input context.
High-quality layouts, rich text and Markdown, which were analyzed by MiniCPM-OCR, were entered directly as context, the core of Agent. It allowed AI agent to remove the first cognitive handicap from the mudslides filled with mudslides and broken formats.
A large number of "unstructured document transfer databases" are just needed in core industries.
For example, tens of thousands of pages of balance sheets, profit statements, invoice base sheets in the area of financial auditing and accounting require a millisecond structured extraction to match entries.
The usual practice is very complex:
Use OCR to convert pictures into text or Markdown, then call the Large Language Model (LLM) with long Prompt and specific JSON Schema rules for "second distillation extraction" and then eventually enter the database.
This has resulted in double VRAM occupancy, high API bills and seconds of delay, and has caused a second hallucination.
The primary key information extracted from MiniCPM-OCR KIE changed this - one step to the end.
Without the need for another large LLM at the back end, key entities can be identified precisely in relation to field and file stream directly to the enterprise-level storage engine.
It's a good time to look at this line in a larger industry.
The global OCR market is still growing rapidly. According to The Business Research Company, market size will increase from $16.66 billion in 2025 to $19.2 billion in 2026 and is expected to reach $33.92 billion in 2030.
But the logic of growth has changed.
Businesses used to pay because "more documents need to be digitized"; now, businesses want digitized results that are "trustable and automated." Financial insurance and health care combined contribute more than 40 per cent of the market share to the wrong zero tolerance industry.
Meanwhile, "transparency" and "auditability" are moving from subdivisions to access thresholds. An analysis model that will only output results and will not answer "is the result directly usable"? A calibrated decision-making layer that says "I'm not sure" is the next stage.
When everyone was still rolling the accuracy rate, the wall had already reached the next road. More importantly, this time, "credible" is open.
One of the most remarkable innovations of human civilization is not the perfect evolution of one person, but the establishment of a system of mutual supervision and calibration.
Now, it's AI's turn to civilization.
Let the invisibility be verified and the trust manifest.
Editor: David Peaches Moses
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.