Liquid AI, founded in 2023 by former MIT researchers, released LFM2.5-2.6B this week: a 2.6 billion parameter, open-weight language model that runs on CPUs without cloud or GPU dependencies. It fits in under 2.5 GB of memory, delivers roughly 220 tokens per second on an Apple M5 Max, 30 tokens per second on a smartphone, and runs on a Raspberry Pi. It supports a 128,000-token context window with native tool calling, and is available now on Hugging Face with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX.
The model was not trained to compete with frontier models. It was trained to replace cloud calls in regulated industries, air-gapped environments, vehicles, and robotics, where latency, privacy, or connectivity make cloud inference impractical. The training pipeline is the real story here: 34 trillion pretraining tokens, a 128K vocabulary, and a four-stage post-training process involving supervised fine-tuning, domain expert distillation, multi-domain on-policy distillation, and reinforcement learning inside production agent harnesses including Hermes Agent and OpenClaw. That last stage produced unexpected gains in math, instruction following, and code.
Liquid also built its own phone-native agent harness, the only one of its kind according to its head of post-training Maxime Labonne, designed for proactive background agents that monitor calendars and trigger tasks without user prompts. Read the full piece for the technical breakdown of the MOPD training stage, the licensing caveats enterprise legal teams need to review, and why Liquid sees the harness as a reliability layer that compensates for model weaknesses rather than a marketing feature.
[READ ORIGINAL →]