Releasing LFM2.5-8B-A1B, a model for local agents

Added: 2026-06-03

[00:00:00]Today we are releasing our LFM 2.5 8B mixture of experts model. This model is made for one specific purpose local agentic workflows.

[00:00:13]The model performs state-of-the-art in many benchmarks that are relevant for agentic use cases such as the tow bench telecom that we have here the Berkeley function calling uh eval as well as multi-turn instruction following benchmark.

[00:00:29]This model is based on the LFM2 mix of expert architecture which is extremely fast and optimized for CPU inference. So on your laptop you can run this model with like hundreds of tokens per second.

[00:00:40]Interesting for this release is that we extended the tokenizer of these models to support non-Latin languages much better. For instance in Hindi we saw 120% increase and improvement in token efficiency of our tokenizer.

[00:00:55]We have zero day support for these models with many popular inference engines including our own leap uh edge SDK Llama CPP MLX VLM sang for hosted inference and many others. We see the model performs topnotch in terms of speed on AMD devices on Apple devices on Qualcomm devices as as well as hosted in the cloud using SG lang as inference engine. If you have aic use cases that require high volume low cost this model might be for you. We also have an open-source desktop app called local cowwork where you can check out this model in action. As always, the model is open on hugging face available for everybody.