on-device AI

Liquid AI has published GGUF builds of LFM2.5-24B-A2B, a 24B-parameter model on the lfm2moe architecture, in sizes from 13.6 GB (Q4_0) to 48 GB (BF16). The Q4_0 file is a Quantization-Aware Distillation checkpoint trained to recover accuracy lost to 4-bit quantization.

Liquid AI released LFM2.5-24B-A2B, an open-weight reasoning mixture-of-experts model with 24B total and 2B active parameters and a 131K-token context window. The company claims it is competitive with Qwen3.6-35B-A3B on agentic and tool-use tasks while decoding at 184 tokens/s on an Apple M5 Max.

OpenBMB has released MiniCPM-V 4.7 1B, a 1B-parameter multimodal model for single-image, multi-image and video understanding on phones, tablets and local hardware. OpenBMB claims it substantially outperforms Qwen3.5 0.8B in overall score and inference efficiency in Instruct mode. Benchmark figures, context window and training cutoff are not given in the available model card text.

OpenBMB has released MiniCPM-V 4.7 35B-A3B, the first mixture-of-experts model in the MiniCPM-V series, with open weights on Hugging Face. The company claims it runs at 13+ tokens/s using 3 GB of memory on an iPhone 17 Pro Max, and that it beats Qwen3.6 35B-A3B on overall score in less inference time in Instruct mode.

Google has released AI Edge Foresight, a free experimental macOS app that transcribes meetings and audio files entirely on-device. It is optimized for Apple Silicon Macs and uses Google's EmbeddingGemma 2 model, according to The Verge. Google says audio, files, and notes never leave the user's computer.

Microsoft announced an upgrade to Copilot at its Windows and Surface event that gives the assistant access to local files and the ability to take actions across Windows. The company calls the underlying approach "Hybrid Intelligence," which combines local and cloud AI models. Pricing, model details, and availability were not disclosed in the available reporting.

Google announced EmbeddingGemma 2, a 740M-parameter natively multimodal embedding model built on the Gemma 4 architecture and released under Apache 2.0. Google says the quantized model needs about 191MB of active RAM for text-only weights and about 567MB for the full multimodal model on a Pixel 11 Pro. Google also launched a Mac app, AI Edge Foresight, to demonstrate it.

Cactus Compute has released Needle 3, a foundation model compressed to 8-29 MB that runs entirely on-device for tool calling, structured data extraction, and text embedding. The company claims it beats models 10x its size on mobile tool calls while running on hardware as small as microcontrollers.

PrismML released Bonsai 2 27B, a compressed version of Alibaba's Qwen3.8 27B model that shrinks memory footprint by 9x to 10x down to 5.9GB. The startup claims 98% aggregate benchmark parity with the original, up from 95% in its first release, using a ternary weight compression technique.

Perplexity's Mac app now supports Hybrid Compute, which starts tasks in the cloud and shifts sensitive steps to a local model running on-device. The feature requires Apple silicon with at least 24GB of unified memory and uses an open-sourced on-device PII classifier to mask private data before any cloud processing.

Liquid AI released DSpark draft model checkpoints for three LFM2.5 models, enabling speculative decoding that speeds up inference by up to 3.18x on H100 GPUs and 2.87x on-device, with day-one support for llama.cpp and SGLang.

Meta CEO Mark Zuckerberg announced the company will open-source its Muse Spark 1.2 model and launch a new on-device model family called Muse Glimmer. The move positions Meta against closed-model rivals OpenAI and Anthropic and against Chinese open-weight labs like DeepSeek and Alibaba.

Liquid AI has released LFM2.5-2.6B, a 2.6B-parameter model designed to run capable tool-calling agents locally on laptops and phones. The company claims it matches or beats models up to 4x its size on instruction-following and tool-use benchmarks while running under 2.5GB of memory.

Samsung's Galaxy Z Fold 8, Fold 8 Ultra, and Flip 8 are the first devices to ship with Google's Gemini Nano 4 on-device model and the new Gemini Intelligence feature tier. The launch comes with strict hardware requirements including 12GB+ RAM and qualified system-on-chips.

Adobe is testing AI-powered photo editing features in its experimental Project Indigo camera app for iPhone. The AI Playground, available to a small percentage of users for a limited time, includes object removal, style transfers, photo critiques, and custom prompt-based editing powered by Google's Nano Banana model.

Scam.ai announced a partnership with Qualcomm and launched Halo, an on-device deepfake detection model for live video calls. The announcements were made at Computex 2026 in Taipei, where Scam.ai was featured at Qualcomm's booth.

Apple announced five third-generation foundation models at WWDC26, headlined by AFM 3 Core Advanced—a 20-billion-parameter sparse model that runs on-device by activating only 1-4 billion parameters at a time. For the first time, Apple extended Private Cloud Compute to third-party infrastructure, with AFM 3 Cloud Pro running on Nvidia GPUs in Google Cloud.

Google DeepMind has released quantization-aware training (QAT) optimized versions of its Gemma 4 model family in GGUF Q4_0 format. The QAT versions preserve similar quality to bfloat16 while dramatically reducing memory requirements, with models available across the entire Gemma 4 lineup: E2B, E4B, 12B, 26B A4B, and 31B.

Google launched AI Edge Gallery for macOS, allowing Mac users to run Google's Gemma models locally. The platform ships with five Gemma models, including the newly released Gemma 4 12B—a 12-billion-parameter multimodal model that handles text, vision, and audio while running on consumer laptops with 16GB of RAM.

Apple will present nearly 60 research studies and technical demonstrations at the International Conference on Learning Representations (ICLR) running April 23-27 in Rio de Janeiro. Demos include the SHARP model that reconstructs photorealistic 3D scenes from a single image in under one second, running on iPad Pro with M5 chip.