LFM2.5-24B-A2B

Not yet independently verified. Every detail here comes from Liquid AI's own article of 13 October 2026 and its model card; no independent report was found. The benchmark figures quoted are Liquid AI's own evaluation and no score is stored. The material states no hosted API price. We will update this when it can be confirmed, and remove this note.

LFM2.5-24B-A2B is a reasoning model from Liquid AI, published on 13 October 2026 as the successor to LFM2-24B-A2B. It is a mixture of experts model with 24 billion parameters in total and about 2.45 billion active for each token. Each of its 40 layers is built from double-gated short convolution blocks and grouped query attention blocks, and each mixture block holds 64 experts with the top four chosen per token. The context window is 131,072 tokens, the vocabulary is about 128,000 tokens, and the model accepts and returns text only. It always reasons step by step before it answers. The model card lists 16 languages: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian and Polish. Liquid AI aims it at agents that reason, call tools and run on a user's own device, and says it is not suited alone to knowledge-heavy work, where it advises pairing it with retrieval or web search. The weights are open under the LFM 1.0 licence, with Base, FP8, GGUF, ONNX and MLX 4-bit variants listed, and a draft model for speculative decoding marked as coming soon. Liquid AI reports decoding speeds of 184 tokens a second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395. In its own evaluation, run with vLLM against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning, Liquid AI reports 96.75 on IFEval and 80.71 on ToolSandbox for LFM2.5-24B-A2B, but a lower SWE-Bench Verified score than Qwen3.6-35B-A3B, 51.00 against 66.20. These are the vendor's figures and were not independently checked, so no score is stored here. No hosted API price is stated.