LFM2.5-24B-A2B : nouveautés, matériel requis, comment l'utiliser gratuitement

LFM2.5-24B-A2B packs 24 billion parameters but only 2 billion are active per token, making it a fast, private on-device reasoning model. It scores 48.09 on LiveClaw, beating a much larger 35 B-parameter competitor.

LFM2.5-24B-A2B is a text-only reasoning model from LiquidAI. It has a total of 24 billion parameters, but only about 2.45 billion are active for each token thanks to a mixture-of-experts design. The architecture includes 40 layers, with 64 experts per MoE block and a top-4 routing scheme. The model supports many languages, from English to Vietnamese, and is delivered in several formats for different runtimes.

Compared to earlier LFM2.5 models, this version expands the context window from 32 K to 131 K tokens and doubles the vocabulary size from 65 K to 128 K, improving non-Latin tokenization. It is a pure reasoning model that always thinks before answering, inserting a special token at the start of each assistant reply. The mixture-of-experts setup keeps only 2 B active parameters per token, which makes inference cheap and fast on consumer hardware.

According to LiquidAI, LFM2.5-24B-A2B achieves a LiveClaw score of 48.09, outperforming the Qwen3.6-35B-A3B model, which scores 38.19, despite having fewer active parameters. The model runs up to 1.7 times faster than comparable mixture-of-experts models on-device.

The main strength of LFM2.5-24B-A2B is its speed: only 2 B active parameters mean each reasoning token is inexpensive, and the model excels at agentic and tool-use tasks. It handles function calling in a four-step process built into its chat template. However, the model card notes it is not recommended for knowledge-heavy tasks on its own. Pairing it with a retrieval system or web search is advised to compensate for that limitation.

If you run AI models at home or look for free, fast reasoning capabilities, LFM2.5-24B-A2B is a solid choice. It is especially suited for developers building private, on-device agents that need to think before responding and call external tools.

Que faut-il pour le faire tourner chez vous ? En 4 bits, il prend environ 17 gigaoctets de mémoire. La plus petite machine personnelle qui le contient entièrement est RX 7900 XT. Sur RTX 3060 12 GB avec 32 gigaoctets de RAM, il écrit environ 97 tokens par seconde. Sur RTX 4090 avec 64 gigaoctets de RAM, il écrit environ 220 tokens par seconde.

There is no free API for it yet. The weights are open: you can download them from Hugging Face, under the license written in its model card. Arena has not ranked it yet: we add the score as soon as the votes arrive.

Compare LFM2.5 24B A2B with every free model, and check what your own GPU can run, on FreeLLMArena.