Author: RadarAI Editorial
Editor: RadarAI Editorial
Last updated: 2026-09-24
Review status: Editorial review pending
Brief
速报
官方
AI动态
开源
Multi-agent collaboration and open-source models became the most concentrated breakthrough directions this week: Microsoft Research proved that k communicating agents can rival 4k independent agents [2], and the SAT framework from Stanford and Together AI achieved an average accuracy of 66.7% across five benchmarks [7]; meanwhile, Xiaomi open-sourced the omni-modal model MiMo-V2.6 Pro, which benchmarks against Claude Opus 5 and GPT-5.6 Sol on Agent benchmarks [5]...
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
## 🔍 Core Insights
**Multi-agent collaboration** and **open-source models** became the most concentrated breakthrough directions this week: Microsoft Research proved that k communicating agents can rival 4k independent agents [2], and the SAT framework from Stanford and Together AI achieved an average accuracy of **66.7%** across five benchmarks [7]; meanwhile, **Xiaomi** open-sourced the omni-modal model **MiMo-V2.6 Pro**, which benchmarks against **Claude Opus 5** and **GPT-5.6 Sol** on Agent benchmarks [5], and officially announced that **MiMo-V3** will adopt the new **HySparse 2** architecture [12].
## 🚀 Key Updates
- **Tencent WorkBuddy hands-on: turning office Agents into "cyber LEGO"** [0]: Through the Expert Plaza, Security Center, and Team Space, it transforms personal experience into organizationally reusable assets.
- **Microsoft Research: k communicating agents rival 4k independent agents** [2]: Surpasses best@k and known optimal results on ARC-AGI-3 and polyomino packing.
- **Google releases the Gemini 3.8 Flash TTS series of speech models** [3][4]: Supports custom voices in 100+ languages and 2,000+ ready-made voices, and can generate hours of glitch-free audio.
- **Xiaomi open-sources MiMo-V2.6 Pro/Flash omni-modal models** [5]: Pro scores 46 on the Artificial Analysis Intelligence Index, the highest among open-source models.
- **Stanford and Together AI propose SAT self-organizing agent teams** [7]: o3-mini, Claude Sonnet 4, and DeepSeek-V3 team up to learn collaboration strategies, achieving an average accuracy of 66.7%.
- **Luo Fuli officially announces that Xiaomi MiMo-V3 adopts the HySparse 2 architecture** [12]: At 1M tokens, prefill computation is reduced by 5.02x, and KV cache is reduced by 4.5x.
- **Meta AI agent Muse attracts 500,000 users in one week** [19]: Daily active users exceed 250,000, cumulative input exceeds 2 million prompts, tops the App Store, and is simultaneously accused of borrowing from OpenClaw.
- **Anthropic engineer explains why Claude's writing quality has declined** [9]: The new model is optimized for math and code, has learned technical expressions meant for AI to read, and has developed a "Claudeish" writing style.
## 🔗 Sources
[0] WorkBuddy turns office Agents into "cyber LEGO" that anyone can build — https://www.bestblogs.dev/article/5e0c711e05?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[2] Microsoft Research paper studies test-time communication: k-agent teams rival 4k independent agents — https://aihot.news/items/cmue9rxdt0siwrogh7vkadlwl
[3] Google releases Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS speech models — https://aihot.news/items/cmuea0fus0styroghsv6n2nc1
[4] Google DeepMind releases
Multi-agent collaboration and open-source models became the most concentrated breakthrough directions this week: Microsoft Research proved that k communicating agents can rival 4k independent agents [2], and the SAT framework from Stanford and Together AI achieved an average accuracy of 66.7% across five benchmarks [7]; meanwhile, Xiaomi open-sourced the omni-modal model MiMo-V2.6 Pro, which benchmarks against Claude Opus 5 and GPT-5.6 Sol on Agent benchmarks [5], and officially announced that MiMo-V3 will adopt the new HySparse 2 architecture [12].
🚀 Key Updates
- Tencent WorkBuddy hands-on: turning office Agents into "cyber LEGO" [0]: Through the Expert Plaza, Security Center, and Team Space, it transforms personal experience into organizationally reusable assets.
- Microsoft Research: k communicating agents rival 4k independent agents [2]: Surpasses best@k and known optimal results on ARC-AGI-3 and polyomino packing.
- Google releases the Gemini 3.8 Flash TTS series of speech models [3][4]: Supports custom voices in 100+ languages and 2,000+ ready-made voices, and can generate hours of glitch-free audio.
- Xiaomi open-sources MiMo-V2.6 Pro/Flash omni-modal models [5]: Pro scores 46 on the Artificial Analysis Intelligence Index, the highest among open-source models.
- Stanford and Together AI propose SAT self-organizing agent teams [7]: o3-mini, Claude Sonnet 4, and DeepSeek-V3 team up to learn collaboration strategies, achieving an average accuracy of 66.7%.
- Luo Fuli officially announces that Xiaomi MiMo-V3 adopts the HySparse 2 architecture [12]: At 1M tokens, prefill computation is reduced by 5.02x, and KV cache is reduced by 4.5x.
- Meta AI agent Muse attracts 500,000 users in one week [19]: Daily active users exceed 250,000, cumulative input exceeds 2 million prompts, tops the App Store, and is simultaneously accused of borrowing from OpenClaw.
- Anthropic engineer explains why Claude's writing quality has declined [9]: The new model is optimized for math and code, has learned technical expressions meant for AI to read, and has developed a "Claudeish" writing style.
🔗 Sources
[0] WorkBuddy turns office Agents into "cyber LEGO" that anyone can build — https://www.bestblogs.dev/article/5e0c711e05?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[2] Microsoft Research paper studies test-time communication: k-agent teams rival 4k independent agents — https://aihot.news/items/cmue9rxdt0siwrogh7vkadlwl
[3] Google releases Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS speech models — https://aihot.news/items/cmuea0fus0styroghsv6n2nc1
[4] Google DeepMind releases
← Back to Updates