Liquid AI has released the LFM3 model family: six models in the roughly 2B class, all branched from a single 1.7B base. There is one each for text, vision, audio and full multimodality, plus two decision models specialized from the vision and omni versions. The weights are on Hugging Face under an open-weights label, and the models can also be reached through the LFM Playground and Liquid's own API.
Architecturally, LFM3 keeps to Liquid's habit of not building a pure Transformer:
"LFM3 combines three operators: gated short convolutions, sliding-window attention (SWA), and global attention."
Short convolutions and sliding-window attention handle most of the layers, while global attention sits in only a handful of them. That is how the model keeps memory use down at long context lengths.
Taking the scorecard apart
The text model, LFM3-1.7B, posts an average benchmark score of 49.37. Among the models Liquid compares it against, its own previous generation LFM2.5-2.6B scores 31.55, MiniCPM5-2B scores 39.19, and Qwen3.5-9B, which has more than five times the parameters, scores 47.24.
The individual results vary widely. Instruction following is very strong, with 96.66 on IFEval, as is tool calling at 71.39 on BFCL V4, and those are the two abilities on-device agents lean on most. Long context and hard knowledge are where it gives ground: RULER at 128K reaches only 65.56 against 87.11 for Qwen3.5-9B, and GPQA Diamond is 63.37 against 80.40. In other words, the lead in the average rests mainly on instruction and tool-use tasks, and on knowledge-heavy work it is still a 2B model.
The name of the license is not stated in the visible part of the official blog post, so commercial terms should be checked against the model card.
Doing the throughput math
On-device numbers are the headline this time. With llama.cpp, Q4_0 quantization and a 32K context, Liquid says LFM3 decodes 1.8 to 2.8 times faster than the next-fastest model on three devices, with the lowest memory footprint.
There is also a server-side figure: on a single NVIDIA B300, running SGLang in BF16 at 1,024 concurrent requests, the model produces about 3.6 billion tokens per day. A rough calculation: 3.6 billion divided by 86,400 seconds is about 42,000 tokens per second, and spread across 1,024 concurrent streams that is around 40 tokens per second each, well above human reading speed. On that basis, one card could serve more than a thousand conversations at once without stalling, which matters more to teams using small models for bulk classification, extraction and routing.
These figures are Liquid's own measurements and have not been independently reproduced. The prompt and output lengths used in testing are not given either, so results could differ sharply under another workload.
Built for fine-tuning
The selling point Liquid stresses again and again is that every model can be fine-tuned on a single GPU. Because the six models share one base, a team can first tune business capability on the text version and then carry it over to the vision or audio branch. Liquid also provides documentation for running the models locally and for fine-tuning them.
For phones, in-car systems and industrial controllers, where a network connection is inconvenient, a 1GB footprint, a 256K context and single-card tunability are the niche LFM3 is trying to claim. Whether it gets there will depend on how the community's fine-tuning results hold up on real tasks, not just on this scorecard.
Sources: Liquid AI official blog, CocoLoop, Hugging Face model page. Benchmark scores, the on-device decoding multiples and the B300 daily throughput all follow Liquid's own measurement basis; the per-second throughput is an editorial back-of-the-envelope figure derived from the daily number.