DeepSeek Open-Source Visual Reasoning Framework; USTC & Huawei Launch Lingjing Zaowu Multi-Agent Platform · 0502-255

2026-05-02 00:00

Author: RadarAI Editorial Editor: RadarAI Editorial Last updated: 2026-08-12 Review status: Editorial review pending Brief 速报官方 AI动态开源

Multimodal reasoning and multi-agent collaboration are emerging as dual technical frontiers: DeepSeek open-sourced a vision-based reasoning framework to bridge spatial reference gaps; USTC and Huawei launched the 'Lingjing Zaowu' platform, enabling autonomous task division and closed-loop execution via Coordination Engineering.

Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.

## 🔍 Key Insights **Multimodal reasoning** and **agent collaboration architectures** are emerging as the two dominant technical trajectories: DeepSeek has open-sourced a visual primitive reasoning framework that directly bridges the spatial reference gap; meanwhile, the “Lingjing Zaowu” platform—jointly launched by USTC and Huawei—leverages a full-stack **Coordination Engineering** system to enable autonomous task division and closed-loop execution among multiple agents [2][8]. ## 🚀 Key Updates - **DeepSeek open-sources new multimodal technology** [8]: Introduces a visual primitive reasoning framework where bounding boxes and coordinate points serve as the smallest units of reasoning—matching the performance of cutting-edge models like GPT-5.4 - **Huawei and USTC launch “Lingjing Zaowu,” an intelligent scientific research cloud platform** [2]: Built on the openJiuwen community’s full-stack **Coordination Engineering**, it supports autonomous multi-agent collaboration and end-to-end scientific workflows - **SkillClaw open-source project launches** [4]: Automates AI agent skill extraction, optimization, and cross-device sharing—with built-in collective evolution mechanisms - **Pentest Swarm AI, an open-source penetration testing tool, is released** [12]: Uses swarm intelligence architecture—multiple AI agents collaborate via pheromone-weighted signals and a shared dashboard to execute full-cycle penetration tests - **Eywa paper proposes a novel co-execution paradigm for LLMs and domain-specific models** [9]: Builds bidirectional interfaces to bridge LLMs and scientific models—resolving the “LLMs don’t understand data; domain models can’t speak” dilemma - **Zhipu publicly analyzes root causes and fixes for GLM-5’s “degradation” issues** [10]: Identifies flawed inference state management under high load as the source of garbled output and repetition—and shares key engineering lessons from large-scale model scaling - **Codex CLI adds `codex features enable goals` command** [1]: Enables goal-based functionality instantly—no config file edits required—dramatically lowering the barrier to entry - **Demis Hassabis predicts AI’s next frontier: virtual cells** [13]: In a YC interview, he notes agents remain in early stages—and AGI evolution will shift toward embodied, simulatable *microscopic* intelligent systems ## 🔗 Sources [1] A More Convenient Way to Launch Codex CLI — https://www.bestblogs.dev/status/2050226611434234358?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [2] Huawei and USTC Launch Lingjing Zaowu; Coordination Engineering Debuts on openJiuwen — https://www.bestblogs.dev/article/5bff5652?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [4] SkillClaw: An Open-Source Project for Automated Evolution and Accumulation of AI Agent Skills — https://www.bestblogs.dev/status/2050206139598721028?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [8] Just Now: DeepSeek Open-Sources New Multimodal Technology!! — https://www.bestblogs.dev/article/7c937a30?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [9] Eywa: Bridging Language Models and Domain-Specific Models to Solve Scientific AI Challenges — https://www.bestblogs.dev/status/2050185392431755366?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [10]

Multimodal reasoning and agent collaboration architectures are emerging as the two dominant technical trajectories: DeepSeek has open-sourced a visual primitive reasoning framework that directly bridges the spatial reference gap; meanwhile, the “Lingjing Zaowu” platform—jointly launched by USTC and Huawei—leverages a full-stack Coordination Engineering system to enable autonomous task division and closed-loop execution among multiple agents [2][8].

🚀 Key Updates

DeepSeek open-sources new multimodal technology [8]: Introduces a visual primitive reasoning framework where bounding boxes and coordinate points serve as the smallest units of reasoning—matching the performance of cutting-edge models like GPT-5.4
Huawei and USTC launch “Lingjing Zaowu,” an intelligent scientific research cloud platform [2]: Built on the openJiuwen community’s full-stack Coordination Engineering, it supports autonomous multi-agent collaboration and end-to-end scientific workflows
SkillClaw open-source project launches [4]: Automates AI agent skill extraction, optimization, and cross-device sharing—with built-in collective evolution mechanisms
Pentest Swarm AI, an open-source penetration testing tool, is released [12]: Uses swarm intelligence architecture—multiple AI agents collaborate via pheromone-weighted signals and a shared dashboard to execute full-cycle penetration tests
Eywa paper proposes a novel co-execution paradigm for LLMs and domain-specific models [9]: Builds bidirectional interfaces to bridge LLMs and scientific models—resolving the “LLMs don’t understand data; domain models can’t speak” dilemma
Zhipu publicly analyzes root causes and fixes for GLM-5’s “degradation” issues [10]: Identifies flawed inference state management under high load as the source of garbled output and repetition—and shares key engineering lessons from large-scale model scaling
Codex CLI adds codex features enable goals command [1]: Enables goal-based functionality instantly—no config file edits required—dramatically lowering the barrier to entry
Demis Hassabis predicts AI’s next frontier: virtual cells [13]: In a YC interview, he notes agents remain in early stages—and AGI evolution will shift toward embodied, simulatable microscopic intelligent systems

🔗 Sources

[1] A More Convenient Way to Launch Codex CLI — https://www.bestblogs.dev/status/2050226611434234358?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[2] Huawei and USTC Launch Lingjing Zaowu; Coordination Engineering Debuts on openJiuwen — https://www.bestblogs.dev/article/5bff5652?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[4] SkillClaw: An Open-Source Project for Automated Evolution and Accumulation of AI Agent Skills — https://www.bestblogs.dev/status/2050206139598721028?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[8] Just Now: DeepSeek Open-Sources New Multimodal Technology!! — https://www.bestblogs.dev/article/7c937a30?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[9] Eywa: Bridging Language Models and Domain-Specific Models to Solve Scientific AI Challenges — https://www.bestblogs.dev/status/2050185392431755366?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[10]

← Back to Updates