What are LFM2.5-Encoders?
LFM2.5-Encoders are new encoder models that match the quality of larger models but remain fast even with long inputs, allowing for document-scale jobs on existing hardware, including CPUs. They can be used for various NLP applications such as intent routers, policy linters, and text classifiers.
The arrival of LFM2.5-Encoders is a big deal in the natural language processing (NLP) world - these two new encoder models are packing a punch despite their relatively small size. Take LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, for instance: they're perfect for all sorts of applications, from intent routers and policy linters to text classifiers. What's really cool about these encoders is how they can handle super long inputs - we're talking up to 8,192 tokens - without slowing down, whether you're running them on a CPU or GPU.
The secret to their success lies in how they were built: a mix of bidirectional attention masks, non-causal short convolutions, and masked language modeling. It's a two-stage training process, first on a short-context masked-language objective and then on a long-context adaptation task - this helps them develop some serious language chops and adapt to all sorts of tasks. When you look at the benchmark results, it's clear these encoders can hold their own against larger models, even beating them on tasks like GLUE, SuperGLUE, and multilingual classification.
One thing that really stands out is how fast these encoders can run - especially on CPU, where they leave ModernBERT-base, a bigger model, in the dust by a factor of 3.7. This means devs can use them to scan or classify long docs, like contracts or transcripts, in under 30 seconds on just a laptop CPU. The demos show off some pretty cool potential applications, from zero-shot prompt routing and policy linting to spell checking and PII detection.
All in all, LFM2.5-Encoders are a game-changer for NLP applications - they strike a great balance between model size, performance, and inference speed. Plus, they can be fine-tuned for specific tasks and used in all sorts of apps, making them a fantastic addition to the NLP community.
Source: Hugging Face
Related posts
Apple Adds AI Assistant to Pages Keynote Numbers
Apple is rolling out a new AI assistant across Pages, Keynote, and Numbers that can design slides, draft text, restructure documents, calculate results, and visualize data by analyzing on-screen content and applying changes directly. The feature launches in beta for devices running iOS 27, iPadOS 27, or macOS 27 and later, using models from Apple and OpenAI. It also adds Model Context Protocol support for third-party AI agents.
DeepMind Unveils Genie 4 World Model With Better Physics and Audio
Google DeepMind unveiled Genie 4, a world model generating interactive simulated environments from text descriptions. Powered by Gemini 3.8 Flash for reasoning, it features improved physics, realistic human movement, and native audio generation-including adaptive spoken dialogue-a first for DeepMind. The simulation remains consistent for only a few minutes. Access is limited to trusted testers, developers via Gemini API and AI Studio, and Google AI Ultra subscribers ($100/month), with
Google DeepMind Unveils Genie 4 World Model With Native Audio
Google DeepMind unveiled Genie 4 on October 27, 2026, a natively multimodal world model generating audio, music, ambient noise, and dialogue, simultaneously with visual content in real-time interactive simulations. Built on the Gemini model family, it maintains spatial and temporal coherence across extended sessions. The research preview is limited to trusted testers, Google AI Ultra subscribers in Project Genie, and early partners, with a public waitlist open but no timeline for broad