24.0B total / 2.45B active (MoE)
131,072 tokens
LFM Open License v1.0
Same-day open-weight release of base and post-trained text-only checkpoints for Liquid AI's largest and final LFM2.5 model. Its sparse hybrid architecture contains 40 layers comprising 30 double-gated short-convolution blocks and 10 grouped-query-attention blocks, with 64 experts per MoE block and top-4 routing. The release doubles the predecessor's vocabulary from 65K to 128K tokens, extends context from 32,768 to 131,072 tokens, supports 16 languages, and reports a 36T-token training budget: approximately 33T pretraining tokens followed by 2T tokens of curated 32K-context mid-training and 300B tokens at 131K context. The post-trained checkpoint is a reasoning-only model whose chat template always begins responses with an explicit chain of thought. Its five-stage post-training pipeline combines supervised fine-tuning, teacher specialization, non-agentic and agentic Multi-Domain On-Policy Distillation, and agentic reinforcement learning inside OpenClaw, Hermes Agent, Pi, and OpenCode. Liquid AI reports 48.09 on LiveClaw, 27.69 on AutomationBench, 65.01 on BFCLv4, 80.71 on ToolSandbox, 96.75 on IFEval, and 51.00 on SWE-bench Verified using OpenCode. Reported deployment performance reaches 184 tokens per second on an Apple M5 Max, 114 tokens per second on an AMD Ryzen AI Max+ 395, and more than 21K aggregate output tokens per second on one NVIDIA B300 at high concurrency. Official deployment formats include native BF16, FP8, GGUF, ONNX, and MLX checkpoints with support for Transformers, vLLM, SGLang, llama.cpp, and Apple Silicon.