Image created with OpenAI GPT-Image-1. Image prompt: Cheesy late-night infomercial freeze-frame—news-desk “Action 6” lower-third featuring map pin product “LOCAL AI LO-PRO™”; headline ticker, CRT curve, high-resolution
With Apple Intelligence, @Apple may have triggered a major realignment in how agentic AI is built — shifting it onto devices. Just look at what we have: ▪️ By opening its on-device model to developers, Apple enables a new gen of apps and models to live in the OS — no cloud https://x.com/TheTuringPost/status/1935470371538645491
Apple refreshed its Apple Foundation Models (AFM) with new versions for on-device and server use, aiming to improve performance in tasks like image understanding and multilingual reasoning. The company also released a Foundation Models API for developers to integrate the https://x.com/DeepLearningAI/status/1936121879552537056
We benchmarked Apple’s new On-Device model: trails most Gemma and Qwen on-device suitable models but still very useful GPQA Diamond performance trailed models that are suitable for on-device use such as the smaller Gemma models (3n E4B, 4B, 12B) and Qwen3 models (1.7B, 4B, 8B). https://x.com/ArtificialAnlys/status/1936141541023924503
Browsers are the perfect place for hybrid AI inference, combining the power of cloud AI with the versatility and privacy of local models. ⚡️ Dia already has great WebGPU support, so I’m looking forward to seeing more on-device models being integrated directly into the browser https://x.com/xenovacom/status/1935092938922758481
Built an internal tool to connect any remotely hosted MCP server and test it against multiple agent architectures to perform evals. Just paste the MCP server URL and hit connect. Database is SQLite and all agent run logs are stored locally. https://x.com/trillhause_/status/1931525724713705589
I wonder if Apple’s approach (on-device LLM with LoRAs & stateless cloud LLM to preserve privacy) makes sense in 2026. It was an alright bet last year, but that isn’t where the market has gone. We got a world with complex multimodal chatbots that people form ongoing bonds with.”” / X https://x.com/emollick/status/1933565955092746483
Wait I forgot every single person at OAI has to love macs and use exclusively macs – so it actually could be huge since macs actually have useable memory”” / X https://x.com/Teknium1/status/1934932729801658661
Gemma 3n is the first model with less than 10B parameters with a LMArena score above 1300 🔥 And yes, you can run it in your phone Try it in AIS https://x.com/osanseviero/status/1934545142393737460
Nice improvement to the Hugging Face hub – you can now filter models with a given size range that run in mlx / mlx-lm: https://x.com/awnihannun/status/1934655784547439008
I woudlnt even think its 30b if it can run “at home” most people are running 12gb cards or less – this thing is going to be tiny I think”” / X https://x.com/Teknium1/status/1934853514645373270
Mistral Small 3.2 is all you need, Apache weights btw”” / X https://x.com/qtnx_/status/1936093789442973902
RT @MistralAI: Introducing Mistral Small 3.2, a small update to Mistral Small 3.1 to improve: – Instruction following: Small 3.2 is bett…”” / X https://x.com/GuillaumeLample/status/1936104812447514968
RT @iScienceLuvr: This is pretty cool, a DeepSeek researcher open-sourced “”nano-vLLM””, a lightweight implementation of vLLM in ~1,200 lines…”” / X https://x.com/jeremyphoward/status/1935994549882830993
