​Doing More With Less: MatFormer and Kimi K2.5

​Description: Two papers exploring how to get more out of modern models. MatFormer nests smaller models inside larger ones, powering Gemma 3n's on-device multimodal AI. Kimi K2.5 combines long-context vision with refined training recipes to build models that see, reason, and act across tasks. Together they ask: how do we make models more capable without just making them bigger?

​Topics

1. ​MatFormer - Nested Transformer for Elastic Inference: Introduces a nested Matryoshka Transformer where a single trained model contains smaller fully functional versions of itself. The architecture behind Gemma 3n's ability to run frontier-level multimodal AI in 2GB on a phone. Raises big questions about whether we need separate model sizes at all.
   https://arxiv.org/abs/2310.07707
   
   

2. ​Kimi K2.5 - Visual Agentic Intelligence Is the Next Step: moves beyond text-only agents toward systems that see, reason, and act in visual worlds. A useful reference for where multimodal agents are heading.
   https://arxiv.org/pdf/2602.02276
   

​Thank you Sky9 for Capital for offering us their office!

​We are excited to have Bernett from Google chat with us about MatFormer.

​Bernett Orlando is a Senior Software Engineer at Google working at the intersection of AI systems, mobile hardware, and real-time multimodal intelligence. His work focuses on bringing cutting-edge research models into practical, on-device experiences—making AI agents faster, more efficient, and accessible directly on edge devices. At Google, Bernett works on Gemini Nano and Project Astra, optimizing multimodal models for low-latency visual agent systems on Pixel devices. His efforts center on reducing Time-To-First-Token (TTFT) and enabling real-time, offline AI capabilities, helping bridge the gap between large-scale research models and the constraints of mobile hardware.

​Beyond engineering, Bernett is passionate about education and community impact. He runs a YouTube channel (@CSinTamil) with 100K+ subscribers, focused on making computer science more accessible to the Tamil-speaking community. He is also a former world champion in blindfold speedcubing (5×5).

​The Organizers:

​AI+

​AI+ is the premier community of 50,000+ AI founders, researchers, and builders. Founded in the San Francisco Bay Area, AI+ has held over 200 events and helps to drive adoption for some of the best AI companies worldwide. 

​AI+ also organizes AI+ Renaissance, the headline AI conference in San Francisco, with speakers including founders and executives from Parallel Web Systems, Wispr Flow, Replit, Windsurf, Neo4j, MindsDB, and 40 leading figures in AI.

​Sky9 Capital

​Sky9 Capital is a global VC with ~$2B AUM investing from early to growth stage in technically driven founders building category-defining companies. Sky9 has 35+ exits/IPOs with a strong focus on AI infrastructure, hardware, and deep tech.

​Registered in Singapore with presence in the Middle East, North America and Asia, Sky9 has backed 150+ companies including TikTok, Moonshot AI, Rox Motors, and Webull, and acts as a long-term partner providing technical insight, global market access, and scaling support.

​Whitepaper Reading Club

​Whitepaper Reading Club is a community of 700 builders, founders, researchers, in Singapore, Malaysia, San Francisco, Bangkok, New York and now Lagos. We come together in person every month to discuss the latest blockchain projects (website)

​​​​​​​Bittensor - discussion (Jul 2025)

​​AlpenGlow (Aug 2025)

​Z Potentials

​Z Potentials - A next-generation founder and tech community built around the intersection of AI, frontier technology, and global innovation. With over 1 million followers and 500+ active founder communities, we have hosted 50+ invitation only events and panels with top-tier venture capital firms.