Liquid AI releases LFM3 family of six open-weight models on one backbone

Liquid AI has released LFM3, a family of six open-weight models for text, vision, audio, omni and decision tasks. All six branch from one base model, LFM3-1.7B-Base, according to a company blog post dated October 27, 2026. Four of the models cover input types: text, vision, audio and omni. The two decision models are specialized from the vision and omni models. All are in the roughly 2B-parameter range. The base and post-trained weights are available on Hugging Face, Liquid's Playground and its API. Liquid says every model can be fine-tuned on a single GPU.

The architecture mixes three operator types. Of the 24 layers in LFM3-1.7B, 13 are gated short convolutions, 5 are sliding-window attention with a 256-token window, and 6 are global attention. The sliding-window layers look back only 256 tokens. The company says a new tri-hybrid architecture and training recipe let the 1.7B model outperform its previous LFM2.5-2.6B while using 35% fewer parameters. Pretraining used 22T tokens, and the final 3.5T of those were a mid-training stage that also extended the context window to 256K.

The text model fits in about 1 GB at 4-bit precision. On-device numbers were measured with llama.cpp at Q4_0 and a 32K-token context. The post says LFM3-1.7B decodes 1.8 to 2.8 times faster than the next-fastest model on three devices. At a full 128K context, its KV cache needs 1.6 GB, against 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B.

The benchmark tables are Liquid's own. On agentic tasks, LFM3-1.7B scores 38.42 on WebArena-Verified against 28.94 for Qwen3.5-9B, and 71.39 on BFCL V4 against 62.83. It trails the same Qwen model on GPQA Diamond, at 63.37 against 80.40, and on SWE-bench Verified, at 43.40 against 51.20. The post concedes that larger models still win on agentic coding and knowledge.

The vision model, LFM3-VL-2B, is described as the first LFM vision model that accepts video as input, and it has a step-by-step thinking mode. Across 62 vision benchmarks, its average is 71.4, against 71.7 for Gemma 4 12B and 73.4 for GPT-5.6 Luna. The post uses those averages to support its claim that the model matches Gemma 4 12B and is within 2 points of GPT-5.6 Luna. Its new in-house vision encoder has 164M parameters and, according to Liquid, uses 3 times less compute than the SigLIP2 encoder in earlier LFM vision models.

The audio model, LFM3-Audio-2B, supports 18 languages and uses a fully causal encoder. That lets it process live microphone input and reduces the wait after the speaker stops, with time to first token as low as 50 ms. Its English speech recognition word error rate is 6.2, slightly worse than the 6.1 of LFM2.5-Audio-1.5B in the same table. The post does not comment on that gap. The omni model, LFM3-Omni-2B, is the company's first model to combine vision, audio and language. The excerpt of the post reviewed gives no benchmark figures for the two decision models.

One outside result comes from MacPaw, whose staff AI engineer Ivan Petrukha says LFM3-1.7B beat Qwen3.5-2B on 65 everyday agent tasks in real macOS environments. The reported figures are 64% task success against 48%, and 58% fewer tokens per task. The post does not say how those tasks were chosen or how many runs were made. The result is a single test reported in a quote, on Liquid's own model.