**Introducing LFM2.5: The Next Big Thing in On-Device AI**
Hey everyone, I’m super excited to share with you the latest development in AI – LFM2.5, a revolutionary new family of compact AI models designed for real-world, on-device agents. This cutting-edge tech is built on the powerful LFM2 structure, optimized for lightning-fast and memory-efficient inference on CPUs and NPUs.
What’s all the fuss about, you ask? Well, LFM2.5 comes with five variants – Base, Instruct, Japanese, Vision Language, and Audio Language – all available as open weights on Hugging Face and LEAP. This means you can dive into the code and make the most of it!
**Breaking it Down**
So, what makes LFM2.5 tick? It retains the hybrid LFM2 structure, which is all about scaling the input and output pipeline for fast and efficient inference. And get this – the pretraining for the 1.2 billion parameter spines has been extended from a whopping 10T to 28T tokens! That’s a massive leap forward in performance.
The Instruct variant, in particular, gets a boost with supervised quality tuning, choice alignment, and large-scale multi-stage reinforcement learning focused on instruction following, tool use, math, and data reasoning. This results in top-notch performance in areas like instruction following and data reasoning.
**Text Model Performance that Stands Out**
The LFM2.5-1.2B-Instruct text model has been tested on various benchmarks, and the results are impressive. It achieves an outstanding 38.89 on GPQA and 44.35 on MMLU Professional, outperforming the likes of Llama-3.2-1B Instruct and Gemma-3-1B IT.
**Japanese, Vision, and Audio Variants**
But that’s not all! The LFM2.5 family also includes Japanese-optimized variants, vision language models for multimodal edge workloads, and even audio language models with native speech generation capabilities. These models are designed to handle real-world tasks like document understanding, user interface learning, and multi-picture reasoning under edge constraints.
**What’s Next?**
Ready to dive deeper into the tech and get your hands on the model weights? Check out the [technical details](https://www.liquid.ai/blog/introducing-lfm2-5-the-next-generation-of-on-device-ai) and explore the various models on [Hugging Face and LEAP](https://huggingface.co/collections/LiquidAI/lfm25).
Want to stay up-to-date with the latest AI news and research? Follow us on [Twitter](https://x.com/intent/follow?screen_name=marktechpost), join our [100k+ ML SubReddit](https://www.reddit.com/r/machinelearningnews/), and subscribe to our Newsletter.
**And if you’re on Telegram, we’ve got you covered there too!** Join us on [Telegram](https://t.me/machinelearningresearchnews).
**What’s next for AI?**
We’ve got our eyes on the future, and it’s looking bright! Our latest launch, [ai2025.dev](https://ai2025.dev), is a 2025-focused analytics platform that tracks model launches, benchmarks, and ecosystem activity. It’s a treasure trove of data waiting to be explored!
Stay curious, stay informed, and let’s keep pushing the boundaries of what’s possible with AI!
—
**About the Author**
I’m Asif Razzaq, founder of Marktechpost Media Inc. As a passionate entrepreneur and engineer, I’m dedicated to harnessing the power of Synthetic Intelligence for social good. With over 2 million monthly views on Marktechpost, we’re making AI accessible to everyone, one article at a time.
