Liquid AI

At NYSE with John Furrier, our CEO Ramin Hasani shared how the next chapter of AI will be a hybrid, combining the scale of the cloud with the adaptability of on-device intelligence. Large-scale training happens in the cloud, but privacy-sensitive, latency-sensitive, and efficient applications must stay on-device. These systems need to perform instantly, securely, and without any risk of interruption. This is where highly specialized, compact models thrive. Liquid’s technology was built for this kind of intelligence. From our research at MIT, we developed liquid neural networks, models inspired by biology and designed for adaptability. These systems stay flexible even after learning, able to hold and compress more knowledge within smaller footprints, without sacrificing performance. By drawing from nature, we have built models that deliver powerful intelligence in compact, efficient form, creating the foundation for AI that can run reliably and privately wherever needed. 🎥 Watch Ramin’s full conversation on SiliconANGLE & theCUBE: https://lnkd.in/eBh3qJwU

Liquid AI, AMD, and Robotec.ai have deployed compact foundation models for autonomous agentic robotics:  showcasing a fine-tuned version of our recently released 3-billion parameter Liquid vision-language model (LFM2-VL-3B), running efficiently on AMD Ryzen™, to enable real-time multimodal perception and decision-making at the edge. Using Robotec.ai's RAI framework—a flexible AI agent platform designed for developing and deploying Embodied AI features—the team validated autonomous inspection capabilities through hardware-in-the-loop simulation across hundreds of warehouse scenarios. This approach enabled rigorous testing of the complete system before physical deployment, significantly reducing development time and risk while ensuring robust real-world performance. Check out the demo at the AMD booth at #ROSCon2025 next week with more to follow.

Introducing our new tiny vision model: LFM2-VL-3B 👀 Built for flexibility and performance: > 51.8% on MM-IFEval (instruction following) > 71.4% on RealWorldQA (real-world understanding)  > Excels in single and multi-image understanding and English OCR > Low object hallucination rate (POPE benchmark) > Expanded multilingual visual understanding: English, Japanese, French, Spanish, German, Italian, Portuguese, Arabic, Chinese, Korean This model expands our multimodal portfolio and demonstrates the universal applicability of our hybrid LFM2 backbones. Blog: https://lnkd.in/edFKrAJk HF: https://lnkd.in/eKrJizeA LEAP: https://lnkd.in/ePKp8CRx

Had a great time today chatting about what we do at Liquid AI at the New York Stock Exchange with John Furrier in the NYSE Wired "AI Factories - Data Centers of The Future" Series. Video coming soon! special thanks to Brian J. Baumann, David Vellante, and theCUBE SiliconANGLE Media for the kind invitation, and Gemma Allen for the great support! #NYSE Wired

Enjoy LFM2-1.2B on Qualcomm NPUs in NexaSDK! Special thanks to the Nexa AI for integrating a series of liquid instances.

おはようございます！Liquid Nanosファミリーに新しく追加されたLFM2-350M-PII-Extract-JPを紹介します。  日本語テキストから個人情報(PII)を抽出し、 デバイス上での機密データのマスキングに使える構造化されたJSONを出力します。 データがローカル環境で処理されるため、プライバシーを保護したまま、クラウドモデル並みの精度と速度を実現します。 Huggingfaceモデル: https://lnkd.in/eD3Fi8tR  LEAPでのデプロイ: https://lnkd.in/eum9_3bm Liquidの最新情報はDiscordチャンネルで随時発信しています。ご参加をお待ちしています。https://lnkd.in/eShCjakY

We have a new nano LFM that is on-par with GPT-5 on data extraction with 350M parameters. Introducing LFM2-350M-PII-Extract-JP 🇯🇵 Extracts personally identifiable information (PII) from Japanese text → returns structured JSON for on-device masking of sensitive data. Delivers the accuracy and speed of giant cloud-based models while keeping data where it belongs: fully private on-device. Download on HF: https://lnkd.in/eD3Fi8tR  Deploy with LEAP: https://lnkd.in/eum9_3bm Join our discord channel to get live-updates on latest from Liquid: https://lnkd.in/eShCjakY

Meet LFM2-8B-A1B, our first on-device Mixture-of-Experts (MoE)! 🐘 > LFM2-8B-A1B is the best on-device MoE in terms of both quality and speed. > Performance of a 3B-4B model class, with up to 5x faster inference profile on CPUs and GPUs. > Quantized variants fit comfortably on high-end phones, tablets, and laptops. Enabling fast, private, low-latency applications across modern phones, tablets, laptops, and embedded systems. Quality LFM2-8B-A1B has greater knowledge capacity than competitive models and is trained to provide quality inference across a variety of capabilities. Including: > Knowledge > Instruction following > Mathematics > Language translation Architecture   LFM2-8B-A1B is one of the first models to challenge the common belief that the MoE architecture is not effective at smaller parameter sizes. LFM2‑8B-A1B keeps the LFM2 fast backbone and introduces sparse MoE feed‑forward networks to add representational capacity without significantly increasing the active compute path. > LFM2 Backbone: 18 gated short convolution blocks and 6 GQA blocks.  > Size: 8.3B total parameters, 1.5B active parameters. > MoE placement: With the exception of the first two layers, all layers include an MoE block. The first two layers remain dense for stability purposes. > Expert granularity: 32 experts per MoE block, with top-4 active experts applied per token. This configuration provides a strong quality boost over lower granularity configs while maintaining fast routing and portable kernels. > Router: Normalized sigmoid gating with adaptive routing biases for better load balancing and training dynamics. CPU Performance: Across devices on CPU, LFM2-8B-A1B is considerably faster than the fastest variants of Qwen3-1.7B, IBM Granite 4.0, and others. GPU Performance: In addition to integrating LFM2-8B-A1B on llama.cpp and ExecuTorch to validate inference efficiency on CPU-only devices, we’ve also integrated the model into vLLM to deploy on GPU in both single-request and online batched settings. Our 8B LFM2 MoE model not only outperforms comparable size models on CPU but also excels against those same models on GPU (1xH100) with full CUDA-graph compilation during decode and piecewise CUDA-graph during prefill. Full Blog: https://lnkd.in/ecyCmKHM HF: https://lnkd.in/eKGCJEDk

Today, we expand our LFM2 family to audio. 👂👄 LFM2-Audio is an end-to-end audio-text omni foundation model, and delivers responsive, real-time conversation on-device at just 1.5B parameters. One model. Seamless multimodal support. No chains.  > Speech-to-speech > Speech-to-text > Text-to-speech > Audio classification  > Open weights 10x faster inference vs peers, with quality rivaling systems 10x larger. Efficiency: Efficiency is key for interactive real-time audio scenarios. LFM2-Audio-1.5B has an average end-to-end latency of under 100 ms, highlighting superb efficiency, even faster than models with much fewer than 1.5B parameters. Quality: LFM2-Audio-1.5B performs best-in-class by a large margin on conversational speech-to-speech chat – competitive with larger open models, such as Qwen2.5-Omni-3B (5B), Lyra-Base (9B), and GLM-4-Voice (9B). Model: LFM2-Audio is a novel omni-modal architecture that supports both text AND audio as first-class modalities, in both input and output. On the input side, the model intakes and tokenizes both text tokens and audio tokens into the same latent space. On the output side, the model autoregressively and flexibly generates tokens of either modality, depending on the task. Full Blog: https://lnkd.in/eHdbsAHg HF: https://lnkd.in/eJutwina