Daily AI RoundupsTuesday, 30 June 2026

The 30 Jun 2026 Calathea roundup leads with Meituan open-sourcing LongCat-2.0, a 1.6T-parameter MoE coding model, then moves through UIUC PhDs Yong Yi Bay and Kathleen A. Yearick's arXiv warning that test-time sampling can hit a selection ceiling. It also covers Japan's ¥387B physical-AI foundation-model funding for robotics, Cognition saying Devin Fusion cuts agentic-coding costs by 35%, vLLM v0.24.0 adding MiniMax-M3 support after 571 commits from 256 contributors, Meta's Brain2Qwerty v2 non-invasive brain-to-text decoder in Nature Neuroscience, and two SemiAnalysis reads on Nvidia data-center revenue and Rubin Ultra design changes.

• coding models
• test-time scaling
• robotics
• agentic coding
• model serving
• brain-to-text
• AI infrastructure
• semiconductors

Stories in this roundup

1. 01
   Meituan open-sources LongCat-2.0, a 1.6T MoE coding modelX / Meituan LongCat
2. 02
   UIUC PhDs Yong Yi Bay and Kathleen A. Yearick argue test-time sampling can hit a selection ceilingarXiv
3. 03
   Japan funds ¥387B domestic physical-AI foundation model for roboticsThe Japan Times / Jiji
4. 04
   Devin Fusion: hybrid-model harness cuts agentic coding cost 35%X / Cognition
5. 05
   vLLM v0.24.0 adds MiniMax-M3, 571 commits from 256 contributorsGitHub / vLLM
6. 06
   Brain2Qwerty v2: Meta's non-invasive brain-to-text decoder published in NatureX / Meta AI
7. 07
   SemiAnalysis: Nvidia DC revenue 20% above consensus for 2H FY2027X / SemiAnalysis
8. 08
   Nvidia cancels original 4-die Rubin Ultra, halves size over manufacturing concernsX / SemiAnalysis