Gemma 3n: The developer guide

Introducing Gemma 3n: The developer guide Gemma 3n represents a major advancement for on-device AI, bringing powerful multimodal capabilities to edge devices with performance previously only seen in last year's cloud-based frontier models.

https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide/  

Comments

Post a Comment

Popular posts from this blog

MIT Team Builds a Speech-to-Reality System that Turns Spoken Prompts into Physical Objects Within Minutes

MIT researchers have devloped a system that allows users to speak a request aloud and receive a fabricated object minutes later, demonstrating how natural language, generative AI, and robotics can combine to produce on-demand manufacturing. According to MIT, the work, presented at the ACM Symposium on Computational Fabrication, shows that the system can assemble simple furniture and decorative items from modular parts without requiring users to know 3D modeling or robotic programming. Researchers at MIT’s Center for Bits and Atoms, led by graduate student Alexander Htet Kyaw with collaborators Se Hwan Jeon and Miana Smith, built a workflow that begins with speech recognition and a large language model. The model interprets the user’s request — such as asking for a stool — and passes the result to a 3D generative AI system that produces a digital representation of the object. A voxel-based process then breaks that form into discrete components suitable for robotic assembly. After geomet...
Read more

When to Think Fast and Slow? AMOR: Entropy-Based Metacognitive Gate for Dynamic SSM-Attention Switching Haoran Zheng

Image
 Transformers allocate uniform computation to every position, regardless of difficulty. State Space Models (SSMs) offer efficient alternatives but struggle with precise information retrieval over a long horizon. Inspired by dual-process theories of cognition (Kahneman, 2011), we propose AMOR (Adaptive Metacognitive Output Router), a hybrid architecture that dynamically engages sparse attention only when an SSM backbone is "uncertain"--as measured by prediction entropy. Compared to standard transformers, AMOR gains efficiency by projecting keys and values from SSM hidden states (Ghost KV), reusing the SSM's O(n) computation rather than requiring O(n^2) attention at every layer. On small-scale synthetic retrieval tasks, AMOR outperforms both SSM-only and transformer-only baselines, achieving perfect retrieval accuracy while engaging attention on only 22% of positions. We validate that prediction entropy reliably signals retrieval need, with a gap of 1.09 nats (nearly half t...
Read more