2.69B
131,072 tokens
LFM Open License v1.0
Open-weight dense text-model release comprising the pretrained LFM2.5-2.6B-Base checkpoint and LFM2.5-2.6B, an always-thinking post-trained model designed for agentic workloads, tool use, data extraction, retrieval-augmented generation, and long-context workflows on consumer hardware. The model uses Liquid AI's hybrid LFM2 architecture with 30 layers: 22 double-gated short-convolution blocks and eight grouped-query-attention layers. Its configuration specifies a 2,048-dimensional hidden state, 32 attention heads, eight key-value heads, a 128,000-token vocabulary, and a 131,072-token context window. Liquid AI reports a 34-trillion-token training budget, including a mid-training stage that extended context to 128K. Agentic post-training consists of two supervised fine-tuning rounds, per-domain teacher specialization, multi-domain on-policy distillation, and reinforcement learning conducted inside popular agent harnesses so the model encounters their tools, system prompts, and interaction patterns. The model supports 16 languages, ChatML-style conversations, structured function calling, and explicit reasoning through think tags. Published benchmark results include 51.87 on AIME25, 59.41 on LiveCodeBench v6, 59.17 on IFBench, 80.07 on Multi-IF, 56.88 on BFCLv4, 77.83 on ToolSandbox, and 26.89 on BrowseComp+ with OpenClaw. Liquid AI reports decode speeds of 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395 while using under 2.5 GB of memory, plus nearly 15,000 output tokens per second at high concurrency on a single H100. Official native, GGUF, ONNX, and MLX releases support Transformers, vLLM, SGLang, llama.cpp, MLX, LM Studio, and ONNX Runtime. Liquid AI does not recommend the model for agentic coding or knowledge-heavy tasks.