Liquid AI, the MIT-spinout foundation model startup, released LFM2.5-2.6B on August 4, calling it the flagship of its new LFM2.5 family of small language models built to run AI agents directly on a phone or laptop instead of in the cloud, the company said in its announcement.

A Hybrid Architecture for Local Hardware

The model has 2.69 billion parameters, arranged across 30 layers that combine 22 gated short-convolution blocks with 8 attention blocks — the hybrid design Liquid AI has used since its original LFM2 models. According to the company, it was pretrained on roughly 34 trillion tokens and supports a 128,000-token context window, letting it process long documents or extended agent sessions without losing track of earlier steps.

Liquid AI says the model is built for on-device AI: planning tasks, calling tools, and working through multi-step jobs entirely on local hardware, with no data sent to a server. The company reports LFM2.5-2.6B decodes at 220 tokens per second on an Apple M5 Max chip and 113 tokens per second on an AMD Ryzen AI Max+ 395 processor, while using under 2.5GB of memory; on a phone, it still manages roughly 30 tokens per second. On a single Nvidia H100 GPU at high concurrency, the company reports throughput near 15,000 output tokens per second.

Bigger Sibling to August’s 230M Model

The release follows LFM2.5-230M, a much smaller Liquid AI model aimed at more constrained hardware such as microcontrollers. LFM2.5-2.6B instead targets phones, laptops, and PCs capable of running heavier agentic workloads without a cloud connection.

Open Weights, Company-Run Benchmarks

Both the base and instruction-tuned versions of LFM2.5-2.6B are published on Hugging Face under an open-weight license that Liquid AI says permits unrestricted download, fine-tuning, and deployment. In benchmarks the company ran itself, LFM2.5-2.6B leads Google’s Gemma-4 and Alibaba’s Qwen3.5 models — including variants nearly four times its size — on instruction-following and most tool-use tests, though it trails larger models on coding tasks. Those results come from Liquid AI’s own testing and have not yet been independently verified.

Read also: Download LFM2.5-2.6B on Hugging Face