Liquid AI's third model generation, released October 27, 2026 as a family of roughly 2B-parameter open-weight models that all branch from one pretrained checkpoint, LFM3-1.7B-Base. The base was pretrained on 22T tokens, with the last 3.5T a mid-training stage of curated data with agentic traces and tool calls that also extended the context to 256K. The architecture is what Liquid calls tri-hybrid: 24 layers of 13 gated short convolutions (kernel size 7, up from 3 in LFM2), 5 sliding-window attention layers with a 256-token window, and 6 global attention layers, with head-wise gated attention, partial RoPE on the global layers and sandwich normalization. The 128K vocabulary is shared with LFM2.5-8B-A1B. Liquid says the layout came from an architecture search across 14 device and backend combinations, from phones and a Raspberry Pi 5 to the Qualcomm Hexagon NPU.

The text model, LFM3-1.7B (1.73B), is post-trained for agent harnesses such as Hermes Agent, Pi and OpenClaw, and its headline numbers are agentic. Liquid reports WebArena-Verified 38.42 against Qwen3.5-9B's 28.94, AutomationBench (public) 41.29 against 15.10, BFCL V4 71.39 and SWE-bench Verified 43.40 (OpenCode harness), across a nine-benchmark average of 49.37 against 47.24 for Qwen3.5-9B. It trails the larger Qwen models on knowledge and long context: GPQA Diamond 63.37 against 80.40, RULER 128K 65.56 against 87.11. It needs about 1 GB at 4-bit and, on Liquid's llama.cpp runs at 32K context, decodes 1.8–2.8× faster than the next model compared. All numbers are self-reported, and Artificial Analysis had not listed it at filing.

The rest of the family shares the backbone. LFM3-VL-2B (1.92B) adds a new 164M in-house vision encoder and is the first Liquid vision model to take video. LFM3-Audio-2B adds a causal audio encoder and the LiquidCodec tokenizer for speech in and out. LFM3-Omni-2B (2.17B) combines both, taking text, images, video and audio and answering in text or speech. Two d1 decision models, d1-2B and d1-Omni-2B, are specialized from the VL and Omni models. All ship under the LFM Open License v1.0 with GGUF, MLX, ONNX, FP8 and NVFP4 builds.

Model Details

Architecture DENSE
Parameters 1.73B
Context window 262,144
Training tokens 22T
License LFM Open License v1.0

Benchmark Scores

Benchmark Score Mode
WebArena-Verified 38.42 —
AutomationBench (public) 41.29 —
BFCL V4 71.39 —
SWE-bench Verified (OpenCode) 43.40 —
IFEval 96.66 —
GPQA Diamond 63.37 —
RULER 128K 65.56 —

Variants

Name Parameters Notes
LFM3-1.7B-Base 1.73B Pretrained checkpoint, 22T tokens (last 3.5T mid-training with agentic traces; context to 256K)
LFM3-1.7B 1.73B Post-trained text model for agent harnesses; 256K context
LFM3-VL-2B 1.92B 1.73B LM + 164M vision encoder + 25M projector; text, images and video; 128K context
LFM3-Audio-2B 2B Causal audio encoder + LiquidCodec tokenizer; text or speech in, text and/or 24 kHz speech out; 32K context
LFM3-Omni-2B 2.17B Text, images, video and 16 kHz audio in; text or speech out; 32K context
open-weighton-devicehybrid-architectureagenticmultimodalpretraining

Related