Liquid AI has released LFM3, a family of small open-weight models meant to be fine-tuned for one job and run on the device where that job happens. Every model branches from a single 1.7B base, so a team can pick text, vision, audio, an omni model that does all three, or a decision model, and expect the same core behaviour underneath. For phone apps, factory equipment and robots that cannot rely on a cloud connection, that is the useful part.
The base was pretrained on 22T tokens, the last 3.5T weighted towards agent traces and tool calls, which also stretched the context window to 256K tokens. Its 24 layers mix gated short convolutions with sliding-window and global attention, chosen by a hardware-in-the-loop search across 14 combinations of device and backend, from Raspberry Pi 5 to Qualcomm's Hexagon NPU. Liquid AI says the 1.7B text model beats its own LFM2.5-2.6B with 35% fewer parameters and runs about 25% faster at 8K tokens of context. The clearest gain is on agent tasks: 38.42 on WebArena-Verified, against 3.02 for the older model.
The other models build on that base. LFM3-VL-2B, the first Liquid vision model to take video, raises its average across 62 vision benchmarks from 57.6 to 71.4 with 1.9B parameters, which Liquid AI puts on par with Gemma 4 12B. LFM3-Audio-2B handles speech in 18 languages, and its codec turns tokens into speech about 87 times faster than real time on one CPU thread of an Apple M4 Max. The d1-2B decision model returns calibrated probabilities in a single forward pass and replaces d1-3B. Fine-tuning notebooks cover tasks such as reading receipts into JSON or mapping spoken smart-home commands to function calls.
This continues the line we have reviewed in LFM2.5-24B-A2B and the tiny d1-omni-600m: small models that trade breadth for speed on cheap hardware. The pitch is less about matching frontier models than about being cheap enough to specialise, since every model fits on a single GPU for training.
Caution is warranted. Every benchmark figure comes from Liquid AI, and some comparisons are against its own older models. The release post does not state a licence, and earlier Liquid models shipped under terms that limit commercial use above a revenue threshold, so check each model card before shipping. The robotics results, including 96.2% average success on LIBERO, come from a separate in-house adaptation rather than the released checkpoints.