Liquid AI's LFM2.5-24B-A2B is built for on-device coding agents
The open-weight model stretches context from 32,768 to 131,072 tokens and doubles its vocabulary, aimed at agent harnesses running on a laptop rather than a server.
Liquid AI released LFM2.5-24B-A2B on 13 October, the largest model in its LFM2.5 family and built specifically to work through multi-step agent tasks locally rather than in the cloud. It has 24 billion total parameters with 2.45 billion active per token, and ships as open weights on Hugging Face with a GGUF build for llama.cpp.
The release roughly quadruples context, from 32,768 tokens in the earlier LFM2-24B-A2B to 131,072, and doubles the vocabulary from 65,000 to 128,000 tokens through an in-place tokenizer expansion aimed at non-Latin languages. Liquid says it decodes at 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395.
LFM2-24B-A2B context: 32,768 tokens
LFM2.5-24B-A2B context: 131,072 tokensWhy it matters: this is pitched at agent harnesses, not chat, the kind of multi-step tool-calling loop that coding agents and local automation run. A 24B model with only 2.45B active and day-one llama.cpp support is the category you can actually run on a laptop rather than rent from an API.
The caveat: it ships under Liquid’s LFM Open License v1.0, not Apache or MIT. Commercial use is free only below $10 million in annual company revenue; past that, Liquid requires a separate paid licence.