GitHub

Welcome to the "Ai Daily" section. This is your daily guide for exploring artificially intelligent worlds, and we give you the hot spots in the field of AI every day, focusing on developers, and helping you to understand technological trends and innovative AI applications.

Fresh AI products click on: https://app.aibase.com/zh

MMProLong: A long document LMM training question and answer is far more efficient than OCR

The MMProLong model, published jointly with the Hong Kong Department, made groundbreaking progress in the long file training of the polymode large language model. By optimizing the training data structure, the study significantly enhanced the positioning capabilities of the model in the long context and exceeded the larger open source model in multiple baseline tests.

[AiBase Summary:]

• Enhancement of the efficiency of the processing of long files of multi-modular models through training on questions and answers

The experiment shows that Question and Answer is better than traditional OCR transfer methods.

MMProLong performed excellently in long text stability beyond larger models

The antspoll LingBot-VA thesis was picked up by the robot summit RSS 2026, allowing robots to push and move.

The research paper Causal World Modeling for Robot Control, a collaboration between antspotech and universities such as the Hong Kong University of Science and Technology, was accepted by RSS 2026 of the international robotic top academic conference, marking international recognition of its exploration in the direction of world model-driven robot control.

[AiBase Summary:]

• Through causal world modelling, robots are able to predict environmental changes and generate next action instructions.

LingBot-VA is the first open-source self-return video-action world model with a strong long-term memory.

• LingBot-VA demonstrated excellent data efficiency and generalization in multiple benchmark tests.

For further information: https://arxiv.org/abs/2601.21998

Three, no queues! Queue ma, announce full opening of Copilot and launch of the Knowledge and Skills Sharing Platform

The article introduced the AI Smart Body Product Copilot and launched the Knowledge Skills Sharing Platform. Copilot features memory, personalization and access to the user knowledge base while supporting external models API and custom functional extension. In addition, users can encapsulate workflows as Skill and post them to the Knowledge Square, marking a shift in the ima platform from “knowledge content” management to “knowledge capacity” to “knowledge capacity”.

[AiBase Summary:]

✨ Copilot is fully open and eliminates the queuing mechanism and attracts over 100,000 users to wait.

• Users can seal the workflow as Skill and post it to the Knowledge Square to achieve knowledge-sharing.

The ima platform extends from `knowledge content' management to `knowledge capacity' sedimentation and flow, promoting intellectual development.

4 - The fall in the “crawl-break” policy of the mangrove bottom, with the formal transition to the “cut-off game” of the minibus.

The AID shorts market has seen a big shuffle, the red fruit shorts have eliminated their bottoms, and the industry has moved from traffic raging to content competition.

[AiBase Summary:]

• The abolition of the Ai Myths Floor Policy and the introduction of a “pure split” model for industry.

📉 The ratio of consumption of the Ai comics has increased dramatically, and the number of live short dramas has declined dramatically.

• The National General Directorate of Radio and Television has promoted the development of human short play and has stepped up its support.

5. The first Big 1.58-bit model of China ' s BitCPM-CANN was released by the Qinghua University for Walling Intellectuals

China’s first three-value model, BitCPM-CANN, based on China’s training on the Chinese platform, was released by the Qinghua University of Faces and OpenBMB Open Source Community, marking a major breakthrough in artificial intelligence technology in China. The model has been developed in the field of low-bit training, with a full-strength prototype and performance better than the same size. In addition, all model weights are open, providing developers with innovative tools.

[AiBase Summary:]

China's first large-scale 1.58-bit (BITCPM-CANN) model, based on China's training for the platform, was published.

BitCPM-CANN released about six times the significant dividend during the reasoning phase and could operate on mainstream flagship phones.

All model weights are open and can be accessed by developers through HuggingFace and ModelScope.

Six, xAI Sacrifices "Skiller": 1.5 trillion parameters

xAI has issued a Glock V9-Medium model of 1.5 trillion parameters to enhance its competitiveness on the programming AI track. The model has enhanced the ability to handle complex programming tasks by introducing Cursor ' s code data and is scheduled to be officially released in mid-June.

[AiBase Summary:]

🧠 Model size jumps directly from 0.5T to 1.5T parameters to achieve a qualitative leap.

• Increased capacity to handle complex programming tasks by introducing Cursor code data.

🚀 It is expected to go online in mid-June and will rewrite the baseline battle for the large-scale AI model.

7-Golder's "Questioning the Location of the Shop" Skill Accessed the Sanctuary.

Golde `Query' Skill Access Point, AEEs open shop decision-making, provide a full-process service such as locational recommendations, point assessment and comparison through natural language dialogue, and help entrepreneurs to select their sites efficiently and reduce trial error costs.

[AiBase Summary:]

• The Smart Assistant of the Gauttern site offers a place referral and point assessment service through the Natural Language Dialogue.

This skill is based on high-terrestrial time and space data and industry knowledge, leading to smart decision-making on commercial sites.

Businesses can obtain structured project reports directly through the Nautilus platform in order to improve the efficiency of the business.

8.OpenAIWe built a lab in Singapore and opened our first overseas AI camp.

OpenAI signed a multi-year cooperation agreement with the Government of Singapore for the establishment of the first overseas applied AI laboratory in Singapore, with an investment of over $234 million. The laboratory will focus on the deployment of forward AI, the development of local AI talent and the promotion of AI technology. In addition, the laboratory will create over 200 technical posts for Singapore to help it become a global AI hub.

[AiBase Summary:]

🧠 OpenAI will set up its first overseas applied AI laboratory in Singapore, which will invest more than $234 million.

The laboratory will focus on three core directions: deployment of forward AI, development of local AI talent and promotion of AI technology.

The laboratory will create more than 200 technical posts in Singapore to help it become a global AI hub.