Mistral AI 正式发布 Mistral OCR 3，新模型在光学字符识别领域带来显著进步。价格方面，每1000页处理费为2美元，通过批量API折扣可降至1美元，这一价格优势使其比国内doc2x API更具竞争力，同时 mineru API 虽免费但服务可能有限。在基准测试中，Mistral OCR 3 表现优于 DeepSeekOCR，但对比数据有限，未与 marker、mineru 等同类产品直接比较，未来需更多独立评测。与上一代模型相比，新版本在性能和效率上均有提升，适合批量处理需求。此次发布为 AI 领域的 OCR 技术提供了新的选择，推动了行业竞争与创新。
原文链接：Linux.do
An author on Linux.do conducted an interesting experiment, submitting a matchmaking quiz to multiple AI models. The results showed that Gemini 3 Pro performed best in terms of accuracy and logical reasoning, especially demonstrating exceptional ability in handling counter-intuitive problems. The most challenging question in the quiz was the last one, with the correct answer being "just walk away," an answer that goes against conventional intuition. However, Gemini 3 Pro not only provided the correct answer but also offered an in-depth explanation, revealing the AI's understanding of complex social situations. The author also published a related test questionnaire on Sojump, attracting 6 participants and generating a discussion across 7 posts. This experiment not only showcases the potential of AI in non-traditional problems but also provides a valuable case study for understanding the boundaries of AI reasoning and its practical applications, offering inspiration for readers interested in the forefront of artificial intelligence.
Original Link:Linux.do
This article explores practical deployment challenges for enterprise AI tools, particularly implementation solutions within existing IT security environments. The author plans to replace Codex with ClaudeCode due to better compatibility with company antivirus software, avoiding frequent popup interference issues. The article details the need to build an internal New-API gateway, aiming to allow employees to select from multiple AI models within a secure internal network environment while controlling costs. Current technical challenges include: ClaudeCode not displaying response content from New-API proxy models, and permission error prompts when configuring the Codex monthly subscription channel. This solution offers valuable reference for enterprise technical professionals with similar requirements, covering practical application experience in API integration, balancing enterprise security policies with AI tools.
Original Link:Linux.do
This article conducts a practical performance test of GPT-5.2-Codex. Through a Minecraft clone prompt experiment, it was found that compared to GPT-5.2, there is a significant improvement in processing speed—at least 50% faster, with completion time reduced from about 10 minutes to 4 minutes. However, in terms of final product quality, both achieved similar completion levels with no obvious capability improvements. For comparison, Opus 4.5 was closer to the original textures in visual fidelity but had more bugs, such as interface interaction issues. The test results suggest that current AI model iterations may focus more on efficiency optimization rather than capability leaps, which holds important reference value for AI application development. The author emphasizes that the test is based on subjective evaluation and is for reference only.
Original Link:Linux.do
Proxify, an open-source, lightweight, and self-hosted AI API proxy gateway, has recently released a significant update, adding IP whitelist and Token authentication features, as well as model field mapping capabilities within the request body. The AUTH_IP_WHITELIST can be configured through the .env file to implement IP whitelisting, supporting single IP addresses, CIDR network segments, and multi-rule configurations. Additionally, the routes.json configuration file now includes a model_map parameter, allowing flexible matching and rewriting of the model field in the request body. This high-performance reverse proxy gateway, developed in Go, is specifically designed to address issues such as regional restrictions and complex multi-service configurations, with deep optimization for LLM streaming responses to ensure optimal performance and user experience. Developers can contribute to this practical tool through the GitHub project.
Original link:Linux.do
This article shares tests conducted on the Linux.do forum, evaluating the juice value of the GPT-5.2 model in Codex and Copilot. Test results show that Copilot's spec is medium and cannot access the gpt-5.2 xhigh version. Users explored the model's performance limitations by calculating the juice number divided by 2 multiplied by 10 divided by 5. This finding provides a reference for understanding the practical applications and performance boundaries of AI models, revealing the differences in specifications among AI services.
Original link:Linux.do
A user reports a critical issue with the gcli2 tool during credential acquisition. Previously, credentials worked normally for reverse proxy operations, but subsequently encountered 403 errors. After deleting credentials, connection authentication failed, displaying 'code=not acquired, state=None', and when attempting to obtain credentials, the message 'All connection attempts failed' appeared. The user tried switching multiple nodes, mirror URLs, and official endpoints, but the problem remained unresolved. This help post reveals authentication challenges in tool usage, involving network credential management and connection failure mechanisms, providing practical reference value for other users troubleshooting similar technical issues, reflecting common fault diagnosis needs in technical communities.
Original Link:Linux.do
最新评论
I don't think the title of your article matches the content lol. Just kidding, mainly because I had some doubts after reading the article.
这个AI状态研究很深入，数据量也很大，很有参考价值。
我偶尔阅读 这个旅游网站。激励人心查看路线。
文章内容很有深度，AI模型的发展趋势值得关注。
内容丰富，对未来趋势分析得挺到位的。
Thank you for your sharing. I am worried that I lack creative ideas. It is your article that makes me full of hope. Thank you. But, I have a question, can you help me?
光纤技术真厉害，文章解析得挺透彻的。
文章内容很实用，想了解更多相关技巧。