LFM2.5-2.6B

LFM2.5-2.6B is the text-only flagship of Liquid AI's LFM2.5 family, published on 28 July 2026 and built for deployment on the device rather than in a data centre. It has 2.69 billion parameters across 30 layers, of which 22 are double-gated short-convolution blocks and only 8 use grouped-query attention — the hybrid arrangement that lets the model hold a 131,072-token context without the memory cost of a conventional transformer of the same reach. The model is pre-trained on roughly 34 trillion tokens, with a mid-training phase that stretches the context to 128K, and then put through four post-training stages: two rounds of supervised fine-tuning, per-domain teacher specialisation, multi-domain on-policy distillation and agentic reinforcement learning carried out inside popular agent harnesses, so that the model meets their tools and system prompts during training rather than after release. One detail sets it apart from other small models: it is a pure reasoning model. The chat template inserts a thinking tag at the start of every assistant turn, so it always reasons before answering — there is no toggle to switch that off. Function calling is native, with Pythonic calls between dedicated tokens by default and JSON on request. Sixteen languages are declared, Polish among them. Liquid AI reports 220 tokens per second on an Apple M5 Max and 113 on an AMD Ryzen CPU, inside 2.5 GB of memory, and claims parity with models four times larger on tool use and instruction following. The company is unusually direct about the limits: it recommends the model for agentic workloads, data extraction, RAG and long-context work, and explicitly advises against agentic coding and knowledge-heavy tasks. Weights are downloadable in native, GGUF, ONNX and MLX builds, with a separate base checkpoint for fine-tuning. The licence is not open in the usual sense: the LFM Open License v1.0 permits commercial use only below ten million dollars of annual revenue, above which a separate agreement with Liquid AI is required.

▸Videos

No videos yet.

▸Reviews