LFM2.5 350M
launchearly February 2026
goblin vibe check:
runs on a potato and thinks surprisingly well for something that fits in your phone
compact liquid foundation model tuned for edge reasoning, long streams, and low-latency inference without standard transformer attention costs.
context
infinite
tokens
speed
3x
transformers
350M parameters but competes above its size classInfinite-context style memory scalingEdge-optimized for robots, sensors, and low-power devicesStreaming-friendly low-latency generation
key features
350M parameters but competes above its size classInfinite-context style memory scalingEdge-optimized for robots, sensors, and low-power devicesStreaming-friendly low-latency generation
spec & usage
Built on Linear Dynamical Systems and Structured State Space designs instead of standard transformer attention
Distilled from larger Liquid models to keep reasoning quality on constrained hardware
Compatible with llama.cpp-style local runtimes plus specialized edge inference stacks
limitations
Smaller internal knowledge base than multi-billion-parameter general models
LoRA-style tuning needs Liquid-aware tooling rather than standard transformer adapters
scope:
languageagentresearchlocalopen-sourcefreefastlightweight
launchearly February 2026