LFM 1

Liquid Foundation Models (LFMs) deliver high-performance generative AI through a non-transformer architecture that maximizes throughput while minimizing memory overhead.

Liquid AI (an MIT spinoff) engineered the LFM 1.2B and 2.6B models to solve the scaling and memory bottlenecks of standard transformer architectures. These models deliver top-tier results on benchmarks like MMLU and ARC-Challenge: beating out larger competitors while maintaining a smaller hardware footprint. They utilize linear computational complexity to manage 32k token context windows (ideal for long-form data processing) and run natively on NVIDIA GPUs, AMD hardware, and Apple Silicon. This design ensures fast inference and low VRAM usage for high-density deployment.

https://www.liquid.ai