on-device

PrismML has released Bonsai 27B, claiming it's the first 27-billion parameter model capable of running on-device on iPhone. The model achieves 58-87 tokens per second on Apple's M5 Max chip with a 4GB memory footprint, using 1-bit and ternary quantization to fit within iPhone's approximately 6GB available app memory.

Liquid AI has released LFM2.5-8B-A1B, an 8-billion parameter hybrid model designed specifically for edge AI and on-device deployment. The model is available in multiple GGUF quantized formats ranging from 4-bit (4.84 GB) to 16-bit (16.9 GB), optimized for memory efficiency.

Supertone has released Supertonic 3, a 99M-parameter text-to-speech model that runs entirely on-device using ONNX Runtime. The model expands language support from 5 to 31 languages compared to Supertonic 2, requires no GPU, and claims competitive accuracy against models 7-20x larger.

Google DeepMind released the Gemma 4 E4B assistant model using Multi-Token Prediction (MTP) architecture that accelerates inference by up to 2x through speculative decoding. The 4.5B effective parameter model supports 128K context windows and handles text, image, and audio input with pricing not yet disclosed.

Zyphra has released ZAYA1-8B, a mixture-of-experts language model with 760M active parameters and 8.4B total parameters. The model scores 89.1% on AIME 2026, competitive with models exceeding 100B parameters, while maintaining efficiency for on-device deployment.

Google DeepMind has released Gemma 4, a family of open-weight multimodal models including a 31B dense model with 256K context window and four size variants ranging from 2.3B to 30.7B effective parameters. The release includes Multi-Token Prediction (MTP) draft models that achieve up to 2x decoding speedup through speculative decoding while maintaining identical output quality.

PrismML, a Caltech-founded startup, has released Bonsai 8B, a 1-bit quantized large language model that the company claims is 14x smaller and 5x more energy efficient than full-precision counterparts while remaining competitive with standard 8B models. The model fits into 1.15GB of memory and uses a novel 1-bit weight representation (binary signs with shared scale factors per weight group) instead of traditional 16-bit or 32-bit precision.