Tag: #on-device36 results
Google DeepMind launched Gemma 4, open-weights multimodal models in E2B, E4B, 26B A4B, and 31B sizes. They handle text, images, video, and audio, with up to 256K context and strong reasoning, coding, agentic features. Dense and MoE variants optimized for on-device deployment.
Liquid AI has released LFM2.5-230M, a 230-million-parameter model designed for high-performance data extraction on edge devices. It utilizes the LFM2 architecture to outperform larger models while maintaining a memory footprint under 400MB.
Google's Gemma 4 model is now fully open-source under Apache 2.0. It enables powerful offline multimodal AI on servers, phones, and Raspberry Pi. Developers get total control for edge and on-premises deployments.
A new 1-bit 8B parameter LLM fits in 1.15GB memory and rivals Llama3 8B on benchmarks. It achieves 440 tok/s on RTX 4090, 136 tok/s on M4 Pro, and 40 tok/s on iPhone. The model is available on Hugging Face from a Caltech spinout.
Apple is in early discussions with startup PrismML to utilize technology that shrinks large AI models for on-device execution. This move aims to reduce reliance on cloud processing for features like Siri.
Microsoft has released 'Foundry Local', a platform allowing developers to integrate AI features directly onto user devices. This approach eliminates cloud dependency, reduces network latency, and removes token-based usage costs.
Google is introducing agentic AI and vibe-coded widgets to Android. Gemini Intelligence will feature Gboard-based dictation and form-filling capabilities. This enhances on-device AI interactions.
AI startups shift to small on-device models as cloud LLMs rack up losses like OpenAI's $212B. Optimized 1-7B models like Phi-3, Llama 3.2, MiniCPM rival giants offline on phones. Terminal vendors like Apple lead with NPU integration.
PokeClaw is the first fully on-device app using Gemma 4 to autonomously control Android phones via Accessibility services, with no cloud dependency. It runs entirely on phone CPU using LiteRT, ensuring complete privacy as data never leaves the device. Available on GitHub for one-time model download and offline use.
Demo showcases real-time audio/video input and voice output using Gemma E2B on M3 Pro Mac, ideal for language learning by pointing camera at objects. Model is multilingual, allowing fallback to native languages. Repo provided for replication.