Contents

LFM (Liquid Foundation Models) is the language model line of Liquid AI, the MIT spinoff behind liquid neural networks. The line is notable for building on non-Transformer and hybrid sequence-modeling research rather than the standard Transformer recipe, and for pivoting from proprietary hosted models (LFM1, 2024) to small open-weight models aimed at phones, laptops, and embedded hardware (LFM2 and LFM2.5, 2025-2026).

LFM1 (2024)

Announced on September 30, 2024, the first generation comprised LFM-1.3B, LFM-3B (3.1B parameters), and LFM-40B, a mixture-of-experts model with 40.3 billion total and about 12 billion activated parameters, each with a 32K context window.1 Liquid described the models as built "from first principles" on computational units drawn from dynamical systems, signal processing, and numerical linear algebra, the research lineage that also produced state-space models, though the company disclosed structured-operator design principles rather than a full architecture specification.1 Liquid claimed leading benchmark quality per memory footprint at each size; press coverage noted the claims were developer-reported.2 LFM1 weights were not released: access ran through the hosted Liquid Playground and partner endpoints, with the company stating it was "not open-sourcing our models at the moment."1

LFM2 (2025)

LFM2, released July 10, 2025, reversed the access model: open weights on Hugging Face under the LFM Open License v1.0, an Apache-2.0-based license that is free for research and for commercial use by companies under $10 million in revenue, with larger deployments requiring a commercial agreement.3 The initial checkpoints (350M, 700M, 1.2B, with a 2.6B model following) use a disclosed hybrid architecture: 10 double-gated short-convolution blocks interleaved with 6 grouped-query attention blocks, found via hardware-in-the-loop architecture search under edge latency and memory constraints.3 Liquid reported roughly 2x faster CPU prefill and decode than Qwen3 equivalents, trained on 10 trillion tokens with distillation from an internal LFM1-7B teacher.3 The LFM2-VL vision-language variants (450M and 1.6B, August 2025; a 3B version in October 2025) extended the family to multimodal on-device use.4

LFM2.5 and 2026 releases

The LFM2.5 refresh, announced in January 2026, kept the LFM2 architecture but extended pretraining from 10 to 28 trillion tokens and scaled up reinforcement-learning post-training, shipping open-weight base, instruct, vision, and audio variants.5 Through the first half of 2026 Liquid added, among others, LFM2.5-VL-450M (April), an on-device mixture-of-experts model with 8.3 billion total and 1.5 billion active parameters (May), and a 230M model targeting hardware as small as a Raspberry Pi 5 (June).5

Position

LFM competes on efficiency, not frontier capability: its niche is the on-device segment contested by Phi, Gemma, and small Qwen3 models, and it stands with Jamba and Falcon-H1 among the production bets that post-Transformer and hybrid designs pay off where memory and latency budgets are tightest. Benchmark positioning throughout the line is developer-attributed, as of July 2026.23

See also

References

  1. Liquid AI, "Liquid Foundation Models: Our First Series of Generative AI Models," blog, September 30, 2024. ↩↩↩

  2. SiliconANGLE, "Liquid AI debuts new LFM-based models that seem to outperform most traditional large language models," September 30, 2024. ↩↩

  3. Liquid AI, "Introducing LFM2: The Fastest On-Device Foundation Models on the Market," blog, July 10, 2025; Liquid AI, "LFM2 Technical Report," arXiv:2511.23404. ↩↩↩↩

  4. Liquid AI, "LFM2-VL: Efficient Vision-Language Models," blog, August 2025, and "LFM2-VL-3B: A New Efficient Vision-Language for the Edge," blog, October 2025. ↩

  5. Liquid AI, "Introducing LFM2.5: The Next Generation of On-Device AI," blog, January 2026; MarkTechPost coverage of LFM2.5 releases, January-June 2026. ↩↩