Liquid AI Releases LFM2.5-2.6B: An On-Device Agentic Model with 128K Context, Tool Calling, and Open Weights

Liquid AI released LFM2.5-2.6B, an agentic model that runs entirely on-device, planning, calling tools, and handling multi-step tasks on phones, laptops, PCs, and robots. The model has 2.69B total parameters, a 131,072-token context window, and a 128,000-token vocabulary, pre-trained on approximately 34 trillion tokens. Two checkpoints were shipped: LFM2.5-2.6B-Base for fine-tuning and LFM2.5-2.6B post-trained for agentic workloads. Because inference stays local, data never leaves the device and marginal cost per run is near zero. The model reports tool-use and instruction-following scores competitive with models nearly four times its size. Both checkpoints are public on Hugging Face under the lfm1.0 license, with weights in native, GGUF, MLX, and ONNX formats, and day-one support in llama.cpp, vLLM, SGLang, and LM Studio. The architecture features 30 layers (22 double-gated short convolution blocks plus 8 grouped-query attention blocks) and a vocabulary size of 128,000. Training involved a four-stage post-training process including supervised fine-tuning, teacher specialization with reinforcement learning, multi-domain on-policy distillation, and agentic reinforcement learning with GRPO. Benchmarks show the model leads in all reported instruction-following benchmarks and nearly all tool-use benchmarks, trailing larger models like Qwen3.5-9B only on BFCLv4 and coding tasks. The model is recommended for agentic workloads, tool use, data extraction, RAG, and long-context workflows, but not for agentic coding or knowledge-heavy tasks.