According to the official model card and technical specifications released by Liquid AI on Hugging Face on October 13, 2026, the company has launched LFM2.5-24B-A2B, the flagship and concluding model in its 2.5 foundation series.
The architecture represents a non-traditional hybrid design spanning 40 layers: 30 double-gated short convolution layers designed for linear-time sequential representation, interspersed with 10 grouped-query attention (GQA) blocks for cross-sequence associative retrieval. Operating across 64 experts per feedforward block with top-4 routing, the model maintains a total footprint of 24.0 billion weights while routing only 2.45 billion active parameters per forward pass.
Liquid AI trained the model across a cumulative budget of 36 trillion tokens. Pre-training consumed 33 trillion tokens, followed by a mid-training regime comprising 2 trillion tokens at a 32,768-token context length heavily weighted toward symbolic math, code, and multilingual reasoning. A subsequent 300-billion-token long-context curriculum expanded the effective context window to 131,072 tokens. In-place tokenizer expansion doubled vocabulary capacity from 65,000 to 128,000 tokens to improve non-Latin script compression across 16 supported languages.
Unlike conversational checkpoints that treat reasoning as an optional flag, LFM2.5-24B-A2B enforces mandatory internal reflection. The model automatically emits structured deliberation enclosed within `<think>` tags prior to synthesizing final outputs. On the LiveClaw agentic tool-execution benchmark, the model registered a score of 48.09, outperforming Qwen3.6-35B-A3B (38.19) despite utilizing roughly 50 percent fewer active parameters.
The model is deployed immediately across model gateways, including OpenRouter, where pricing is pegged at $0.05 per million input tokens and $0.70 per million completion tokens. Local deployment formats have been distributed across FP8, ONNX, GGUF, and 4-bit MLX runtimes. Whether the model's specialized convolution blocks can match full dense transformer architectures on open-domain knowledge extraction without retrieval augmentation remains an open empirical question.