Back to the ticker

Liquid AI releases the LFM3 family, led by a 1.7B model for local agents

Liquid AI has released LFM3, a new generation of its Liquid Foundation Models that puts text, vision, audio, omni and decision models on one shared backbone. The first release covers six models in the 2B parameter range, all branched from an LFM3-1.7B-Base checkpoint the company pre-trained on 22 trillion tokens. The text model LFM3-1.7B carries a 256K-token context window, and Liquid AI says it outperforms the earlier LFM2.5-2.6B while using 35% fewer parameters.

Introducing LFM3: Every Modality for Specialized Intelligence

Loading connects your browser to www.youtube-nocookie.com, which may process your IP address and use cookies. Privacy details

Open externally
Video: Liquid AI.

The architecture is a tri-hybrid of 24 layers, 13 of them gated short convolutions, 5 sliding-window attention with a 256-token window and 6 global attention. Short convolutions mix each token with a few of its neighbors instead of attending over the whole history, which keeps the attention cache small. The convolution kernel grew to 7 from 3 in LFM2, a sigmoid gate now scales each attention head’s output, and the global attention layers rotate only the highest-frequency half of each head using partial RoPE. The company redesigned that layout with a hardware-in-the-loop search across 14 device and backend combinations, among them a Samsung Galaxy S26 Ultra, a OnePlus 15, a Raspberry Pi 5, an Apple M4 Max and the Qualcomm Hexagon NPU.

On edge hardware Liquid AI reports the model as the fastest in its own comparison, decoding 1.8 to 2.8 times faster than the next model and using the least memory of the group. Its model card runs that test on Q4_0 GGUF files in llama.cpp with 32,768 tokens of context, on an AMD Ryzen AI Max+ 395, an Apple M5 Max and an iPhone 17 Pro. At 4-bit precision the text model takes about 1 GB of memory, and a full 128K-token context needs 1.6 GB of KV cache against 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B. Executives at AMD, Qualcomm and Ambarella appear in the announcement, with Ambarella saying it has begun enabling LFM3 on the platform that already runs LFM2.5-VL.

The benchmark suite Liquid AI published puts LFM3-1.7B ahead of Qwen3.5-9B on its unweighted average, at 49.37 against 47.24. According to the company the model leads on instruction following and function calling and beats Qwen3.5-9B on long tasks in browsers and business apps, while larger models still win on agentic coding and knowledge. On RULER 128K, which measures recall across a long context, Qwen3.5-9B scores 87.11 to LFM3-1.7B’s 65.56. The model card advises retrieval or search for knowledge-heavy tasks and does not recommend the model for agentic coding or complex multi-step reasoning.

LFM3-1.7B benchmark scores

BenchmarkLFM3-1.7BQwen3.5-9B
LiveClaw45.6035.77
AutomationBench (public)41.2915.10
WebArena-Verified38.4228.94
BFCL V471.3962.83
ToolSandbox80.3676.41
IFEval96.6692.08
SWE-bench Verified (OpenCode)43.4051.20
GPQA Diamond63.3780.40
RULER 128K65.5687.11
Average49.3747.24

All figures according to Liquid AI. The comparison model is Qwen3.5-9B, which has 9.7B parameters against LFM3-1.7B's 1.7B.

Alongside the text model, LFM3-VL-2B pairs the same language stack with a 164M-parameter vision encoder, while LFM3-Audio-2B and LFM3-Omni-2B add a causal audio encoder and the LiquidCodec tokenizer, which packs 80 ms of speech into 12 discrete tokens. The audio model transcribes, synthesizes and translates speech across 18 languages, though its voice chat supports English only for now. Weights ship under the LFM Open License v1.0 in GGUF, MLX, ONNX, FP8 and NVFP4 builds, and the text model covers 16 languages including English, German, Chinese, Japanese and Hindi. Liquid AI runs the family in its own playground and counts two decision models, specialized from the vision and omni models, as part of the same release.

LFM3-Omni-2B: Book a Flight by Voice on Your Laptop

Loading connects your browser to www.youtube-nocookie.com, which may process your IP address and use cookies. Privacy details

Open externally
Video: Liquid AI.
  1. Liquid AI releases LFM2.5-24B-A2B, a 2B-active MoE for local agents
  2. Liquid AI releases d1-3B and d1-omni-600M as open-weight decision models
  3. Liquid AI releases a DSpark drafter that speeds up LFM2.5-VL-3B decoding