Liquid AI has released LFM2.5-2.6B, a 2.69 billion parameter model designed to run entirely on local devices such as phones, laptops, and robots. The model features a 131,072-token context window and was pre-trained on approximately 34 trillion tokens.

  • It supports tool calling and multi-step agentic tasks with weights available in GGUF, MLX, and ONNX formats under the lfm1.0 license.
  • The architecture combines 22 double-gated short convolution blocks with 8 grouped-query attention blocks across 30 layers.
  • Post-training involved four stages including supervised fine-tuning, teacher specialization via reinforcement learning, and agentic reinforcement learning within harnesses like Hermes Agent.
  • Benchmarks show LFM2.5-2.6B outperforming gemma-4-E4B-it and Qwen3.5-9B on ToolSandbox, Multi-IF, and IFStruct, while trailing only on BFCLv4.
  • Inference is highly efficient, decoding at 220 tokens per second on an M5 Max chip using under 2.5 GB of memory.

The open weights allow developers to deploy agentic workloads locally without sending data to third-party APIs, benefiting regulated industries like healthcare and finance.