Streaming experts technology is enabling ultra-large-scale Mixture-of-Experts (MoE) models to run on consumer-grade hardware—demonstrating Qwen with 397B parameters on iPhone and Kimi K2.5 with 1T parameters locally on Mac. Meanwhile, leading AI companies—including Meta, Alibaba, Anthropic, and MiniMax—are accelerating upgrades to agent architectures and advancing the realization of 'Personal Superintelligence' [11][19][24][10][0].
## 🔍 Key Insights
**Streaming experts technology** is enabling ultra-large-scale Mixture-of-Experts (MoE) models to run on consumer-grade hardware—demonstrating **Qwen with 397B parameters on iPhone** and **Kimi K2.5 with 1T parameters running natively on Mac**. Concurrently, industry leaders including **Meta, Alibaba, Anthropic, and MiniMax** are accelerating agent architecture upgrades and the deployment of 'Personal Superintelligence' [11][19][24][10][0].
## 🚀 Major Updates
- **Meta acquires Dreamer to join its Superintelligence Lab** [0]: Accelerating its 'Personal Superintelligence' vision, with focus on desktop AI, self-improving agents, and efficient world models
- **Alibaba launches Accio Work, an enterprise-grade Agent platform** [10]: Enables multi-agent collaboration; real-world testing shows fully automated online store setup and end-to-end operations in **under 30 minutes**
- **Huawei's 'Genius Program' alumnus Kaiwen Zhou joins NoIn Intelligence** [1]: Trains domestic robots using video-generation data; its **KnowinBrain model tops the Embodied Arena leaderboard**
- **Claude's Computer Use feature launched in just 4 weeks** [15]: From Anthropic's acquisition of Vercept to full integration into Claude Code—showcasing exceptional engineering velocity
- **Open-source ChatClaw Agent collaboration system released** [3]: Built atop OpenClaw, supporting multi-company/multi-team/multi-agent architectures and visual configuration
- **MiniMax unveils its M3 multimodal roadmap** [13]: M2.7 demonstrates strong programming capabilities; the next generation will be open-sourced and scale to **1T parameters**
- **SSD-based streaming loading breaks terminal hardware limits** [11][19][24]: Enables 397B Qwen on iPhone and 1T Kimi K2.5 on Mac—redefining the frontier of on-device AI capability
- **OpenAI launches the 'Parameter Golf' challenge** [22]: Inviting developers to submit highly efficient models **≤16 MB**, shifting AI talent evaluation toward capability-driven benchmarks
## 🔗 Sources
[0] [AI News] Dreamer Joins Meta's Superintelligence Lab — A 9-Month Retrospective on Personal Superintelligence — https://www.bestblogs.dev/article/9a259745
[1] Another Huawei Genius Program alum enters embodied AI entrepreneurship! Training home robots with synthetic video data—the first model tops the Embodied Foundation Model leaderboard — https://www.bestblogs.dev/article/60fca1e2
[3] Open-source project ChatClaw: An Agent collaboration system built for OpenClaw — https://www.bestblogs.dev/status/2036316628426826228
[10] Alibaba launches enterprise-grade Agent platform Accio Work overseas: Real-world test completes automated online store setup in 30 minutes — https://www.bestblogs.dev/article/1d9df17b
[11] Streaming Experts — https://www.bestblogs.dev/article/8ad8251d
[13] MiniMax M2.7 Programming Capabilities Demo & M3 Roadmap — https://www.bestblogs.dev/status/2036304982551781678
[15] Anthropic Rapidly Integrates Computer Use Technology — https://www.bestblogs.dev/status/2036302910498173
Streaming experts technology is enabling ultra-large-scale Mixture-of-Experts (MoE) models to run on consumer-grade hardware—demonstrating Qwen with 397B parameters on iPhone and Kimi K2.5 with 1T parameters running natively on Mac. Concurrently, industry leaders including Meta, Alibaba, Anthropic, and MiniMax are accelerating agent architecture upgrades and the deployment of 'Personal Superintelligence' [11][19][24][10][0].
🚀 Major Updates
- Meta acquires Dreamer to join its Superintelligence Lab [0]: Accelerating its 'Personal Superintelligence' vision, with focus on desktop AI, self-improving agents, and efficient world models
- Alibaba launches Accio Work, an enterprise-grade Agent platform [10]: Enables multi-agent collaboration; real-world testing shows fully automated online store setup and end-to-end operations in under 30 minutes
- Huawei's 'Genius Program' alumnus Kaiwen Zhou joins NoIn Intelligence [1]: Trains domestic robots using video-generation data; its KnowinBrain model tops the Embodied Arena leaderboard
- Claude's Computer Use feature launched in just 4 weeks [15]: From Anthropic's acquisition of Vercept to full integration into Claude Code—showcasing exceptional engineering velocity
- Open-source ChatClaw Agent collaboration system released [3]: Built atop OpenClaw, supporting multi-company/multi-team/multi-agent architectures and visual configuration
- MiniMax unveils its M3 multimodal roadmap [13]: M2.7 demonstrates strong programming capabilities; the next generation will be open-sourced and scale to 1T parameters
- SSD-based streaming loading breaks terminal hardware limits [11][19][24]: Enables 397B Qwen on iPhone and 1T Kimi K2.5 on Mac—redefining the frontier of on-device AI capability
- OpenAI launches the 'Parameter Golf' challenge [22]: Inviting developers to submit highly efficient models ≤16 MB, shifting AI talent evaluation toward capability-driven benchmarks
🔗 Sources
[0] [AI News] Dreamer Joins Meta's Superintelligence Lab — A 9-Month Retrospective on Personal Superintelligence — https://www.bestblogs.dev/article/9a259745
[1] Another Huawei Genius Program alum enters embodied AI entrepreneurship! Training home robots with synthetic video data—the first model tops the Embodied Foundation Model leaderboard — https://www.bestblogs.dev/article/60fca1e2
[3] Open-source project ChatClaw: An Agent collaboration system built for OpenClaw — https://www.bestblogs.dev/status/2036316628426826228
[10] Alibaba launches enterprise-grade Agent platform Accio Work overseas: Real-world test completes automated online store setup in 30 minutes — https://www.bestblogs.dev/article/1d9df17b
[11] Streaming Experts — https://www.bestblogs.dev/article/8ad8251d
[13] MiniMax M2.7 Programming Capabilities Demo & M3 Roadmap — https://www.bestblogs.dev/status/2036304982551781678
[15] Anthropic Rapidly Integrates Computer Use Technology — https://www.bestblogs.dev/status/2036302910498173