你的名字

DeepSeek又发新模型，小而美玩出新高度_我的网站

A |     就在刚刚，DeepSeek开源了一个3B模型DeepSeek-OCR。

B |     Sen. Josh Hawley (R-MO) alluded to his Republican colleague’s fentanyl classification bill proposal in a Senate Homeland Security and Governmental Affairs hearing on Tuesday with Gary Rasicot, acting assistant secretary for the US Department of Homeland Security’s Countering Weapons of Mass Destruction Office.,Hawley, who sits on four Senate committees, inquired whether US President Joe Biden’s government intends to declare fentanyl as a weapon of mass destruction in light of the highly fatal and synthetic drug’s growing prevalence in the country.,“Any declaration of fentanyl as a WMD, I think you have to carefully consider that because there are legitimate medical uses for fentanyl,” Raiscot told the Republican lawmaker from Missouri, drawing a distinction between illicit fentanyl and fentanyl cleared for prescriptions.,Fentanyl, also known as ‘China Girl’ or ‘China White,’ is a synthetic opioid estimated to be 50 to 100 times stronger than morphine, according to the DEA. The drug is traditionally used for pain management in cancer patients, but has grown popular among those peddling recreational drugs ranging from marijuana to heroin.,“In regards to illicit fentanyl,” Raiscot noted, “I think you have to proceed with caution there as well because I think what you want to avoid is creating overlapping jurisdictions or even diverting limited WMD resources into a counternarcotics arena.”,Fentanyl can enter one’s body through inhalation, oral exposure or ingestion, or skin contact.,China, once regarded as the primary source of fentanyl in the US, banned the production of fentanyl and substances that mimic the synthetic opioid back in 2019. Nevertheless, US data shows the country remains ravaged by the illicit drug, as over 150 Americans are dying every day from opioids like fentanyl.,Continuing to probe the official from the Countering Weapons of Mass Destruction Office, Hawley raised the possibility of fentanyl being misused by adversaries the US classifies as terrorists.,“Back to this possibility of terrorists of foreign nations weaponizing fentanyl, can you give me your assessment of, is that likely? What are the concerns around that?” Hawley asked.,“I would be happy to give you that assessment,” Raiscot replied, “but I think it would be more appropriately done in a different forum, and I’d like to follow up with you on that.”,Former DEA Agent: Biden’s ‘Lack of Law and Order’ to Blame for 53 Deaths in Texas Lorry Incident2 July, 03:44 GMT,Newly released data from the US Centers for Disease Control and Prevention (CDC) showed overdose deaths in the US have risen 30% within the past year, with “illicitly produced fentanyl” accounting for most of the increase in fatalities.,Florida has emerged as a hotspot for mass-overdose events, prompting a recent plea from the Sunshine State’s attorney general.,“Thinking about curbing the problem in different, new ways may disrupt what the Chinese companies and drug cartels involved are doing or at least make it more expensive or difficult,” Florida Attorney General Ashley Moody said in a memo to President Joe Biden.,The rise in fentanyl-related overdoses and deaths have led many in the DEA to amend their “One Pill Can Kill” campaign slogan.,“Recreational use drugs like cocaine and methamphetamine that are being laced with fentanyl,” DEA Assistant Special Agent in Charge Mike Dubet told local outlet News4Jax. “Therefore, not only can one pill kill you, but one use can also kill you.”,Dubet, who is assigned to the DEA office in Jacksonville, Florida, said Americans are “playing Russian roulette whenever recreational drug use is a part of your life.”。虽然体量不大，但模型思路创新的力度着实不小。                   众所周知，当前所有LLM处理长文本时都面临一个绕不开的困境：计算复杂度是平方级增长的。

C | 序列越长，算力烧得越狠。                             于是，DeepSeek团队想到了一个好办法。既然一张图能包含大量文字信息，而且用的Token还少，那不如直接把文本转成图像？这就是所谓的“光学压缩”——用视觉模态来给文本信息“瘦身”。

D |                    而OCR正好天然适合验证这个思路，因为它本身就是在做“视觉→文本”的转换，而且效果还能量化评估。                             论文显示，DeepSeek-OCR的压缩率能达到10倍，OCR准确率还能保持在97%以上。                   啥意思呢？就是说，原本需要1000个文本Token才能表达的内容，现在只用100个视觉Token就搞定了。即使压缩率拉到20倍，准确率也还有60%左右，整体效果相当能打。

E |                    OmniDocBench基准测试结果显示：                    只用100个视觉Token，就超过了GOT-OCR2.0（每页256个Token）的表现；          用不到800个视觉Token，干翻了MinerU2.0（平均每页超过6000个Token）。                        在实际生产中，一块A100-40G显卡就能每天生成超过20万页的LLM/VLM训练数据。

F | 20个节点（160块A100）直接飙到每天3300万页。

G |                              DeepSeek-OCR由两个核心组件组成：                    DeepEncoder（编码器）：负责图像特征提取和压缩；          DeepSeek3B-MoE（解码器）：负责从压缩后的视觉Token中重建文本。                        让我们来重点说说DeepEncoder这个引擎。                   它的架构很巧妙，通过把SAM-base（8000万参数）和CLIP-large（3亿参数）串联起来，前者负责“窗口注意力”提取视觉特征，后者负责“全局注意力”理解整体信息。                   中间还加了个16×卷积压缩器，在进入全局注意力层之前把Token数量大幅砍掉。

H |                    举例而言，一张1024×1024的图像，会被切成4096个patch token。但经过压缩器处理后，进入全局注意力层的Token数量会大幅减少。

I |                    这样的好处是，既保证了处理高分辨率输入的能力，又控制住了激活内存的开销。

J |                    而且DeepEncoder还支持多分辨率输入，从512×512的Tiny模式（64个Token）到1280×1280的Large模式（400个Token），一个模型全搞定。

K |                    目前开源版本支持的模式包括原生分辨率的Tiny、Small、Base、Large四档，还有动态分辨率的Gundam模式，灵活性拉满。                             解码器用的是DeepSeek-3B-MoE架构。                   别看只有3B参数，但采用了MoE（混合专家）设计——64个专家中激活6个，再加2个共享专家，实际激活参数约5.7亿。这也让模型既有30亿参数模型的表达能力，又保持了5亿参数模型的推理效率。

L |                    解码器的任务就是从压缩后的视觉Token中重建出原始文本，这个过程可以通过OCR风格的训练被紧凑型语言模型有效学习。                   数据方面，DeepSeek团队也是下了血本。                   从互联网收集了3000万页多语言PDF数据，涵盖约100种语言，其中中英文占2500万页。

M |                    数据分两类：粗标注直接用fitz从PDF提取，主要训练少数语言的识别能力；精标注用PP-DocLayout、MinerU、GOT-OCR2.0等模型生成，包含检测与识别交织的高质量数据。                   对于少数语言，团队还搞了个“模型飞轮”机制——先用有跨语言泛化能力的版面分析模型做检测，再用fitz生成的数据训练GOT-OCR2.0，然后用训练好的模型反过来标注更多数据，循环往复最终生成了60万条样本。                   此外还有300万条Word文档数据，主要提升公式识别和HTML表格解析能力。                   场景OCR方面，从LAION和Wukong数据集收集图像，用PaddleOCR标注，中英文各1000万条样本。                             DeepSeek-OCR不仅能识别文字，还具备“深度解析”能力，只需一个统一的提示词，就能对各种复杂图像进行结构化提取：                    图表：金融研究报告中的图表可以直接提取为结构化数据；          化学结构式：识别并转换为SMILES格式；          几何图形：对平面几何图形进行复制和结构化解析；          自然图像：生成密集描述（dense captions）。                        这在STEM领域的应用潜力巨大，尤其是化学、物理、数学等需要处理大量符号和图形的场景。                   第一作者Haoran Wei此前曾供职于阶跃星辰，期间发布并开源了GOT-OCR2.0系统                    值得注意的是，DeepSeek团队在论文里还提出了一个脑洞大开的想法——用光学压缩模拟人类的遗忘机制。                   人类的记忆会随时间衰退，越久远的事情记得越模糊。

N | DeepSeek团队想，那能不能让AI也这样？于是，他们的方案是：                    1. 把超过第k轮的历史对话内容渲染成图像；          2. 初步压缩，实现约10倍的Token减少；          3. 对于更久远的上下文，继续缩小图像尺寸；          4. 随着图像越来越小，内容也越来越模糊，最终达到“文本遗忘”的效果。                   这就很像人类记忆的衰退曲线，近期信息保持高保真度，久远记忆自然淡化。                   虽然这还是个早期研究方向，但如果真能实现，对于处理超长上下文将是个巨大突破——近期上下文保持高分辨率，历史上下文占用更少计算资源，理论上可以支撑“无限上下文”。                             简言之，DeepSeek-OCR表面上是个OCR模型，但实际上是在探索一个更宏大的命题：能否用视觉模态作为LLM文本信息处理的高效压缩媒介？                    初步答案是肯定的，7—20倍的Token压缩能力已经展现出来了。                   当然，团队也承认这只是个开始。

o | 单纯的OCR还不足以完全验证“上下文光学压缩”，后续还计划开展数字–光学文本交替预训练、“大海捞针”式测试，以及其他系统性评估。

p |                    不过不管怎么说，这在VLM和LLM的进化路上，又多了一条新赛道。                   去年这个时候，大家还在卷怎么让模型“记得更多”。今年DeepSeek直接反其道行之，不如让模型学会“忘掉一些”。                   确然，AI的进化，有时候不是做加法，而是做减法。

q | 小而美，也能玩出大花样，DeepSeek-OCR这个3B小模型就是最好的证明。                   GitHub主页：http://github.com/deepseek-ai/DeepSeek-OCR          论文地址：https://github.com/deepseek-ai/DeepSeek-OCR/blob/main/DeepSeek\_OCR\_paper.pdf          模型下载：https://huggingface.co/deepseek-ai/DeepSeek-OCR                    本文来自微信公众号：APPSO （ID：appsolution）。

Current article:http://cq7.yukujuankewazuhangfei.sbs/news/20260826_8551016.html

Published on:09:10:19