Liquid AI Releases LFM3: One 1.7B Base Branches Into Six 2B-Class Open-Weight Models for Text, Vision, Audio, Omni and Decisions, With the Text Model at About 1 GB in 4-Bit

One base, six models

Liquid AI released LFM3, the third generation of its Liquid Foundation Models, on October 27. The first six models are all in the 2B-parameter range and branch from the same LFM3-1.7B-Base. The base was pre-trained on 22 trillion tokens, with the final 3.5 trillion a mid-training stage weighted toward agentic traces and varied tool calls, which also extended the context window to 256K:

• LFM3-1.7B: a text model with 256K context, post-trained to run in agent harnesses such as Hermes Agent, Pi and OpenClaw;
• LFM3-VL-2B: a vision-language model, the first LFM vision model to take video input, with spatial reasoning;
• LFM3-Audio-2B: voice chat, transcription, synthesis and translation across 18 languages;
• LFM3-Omni-2B: vision, audio and language in one model, answering in text, speech or both;
• d1-2B and d1-Omni-2B: decision models specialized from the vision and omni models.

Base and post-trained models are open-weight and available on Hugging Face, as well as through Liquid's Playground and API, and every model can be fine-tuned on a single GPU.

Architecture and official numbers

LFM3 uses what Liquid calls a "tri-hybrid" design: of 24 layers, 13 are gated short convolutions (kernel size up from 3 in LFM2 to 7), 5 are sliding-window attention with a 256-token window, and 6 are global attention. The design came from a hardware-in-the-loop architecture search across 14 device and backend combinations, including the Samsung Galaxy S26 Ultra, a Raspberry Pi 5, an Apple M4 Max and the Qualcomm Hexagon NPU.

Liquid's figures: LFM3-1.7B beats the previous LFM2.5-2.6B with 35% fewer parameters; it fits in about 1 GB at 4-bit; a full 128K context needs 1.6 GB of KV cache, against 4.3 GB for Qwen3.5-4B; and with 8K tokens of context it is about 25% faster than LFM2-1.2B. Liquid also published a comparison on 65 real macOS agent tasks: a 64% success rate for LFM3-1.7B versus 48% for Qwen3.5-2B, with 58% fewer tokens per task and getting stuck in under 1% of attempts (versus 15%); an engineer at Mac software company MacPaw endorsed the result in the announcement. The vision models' new encoder uses a third of the compute of the SigLIP2 encoder it replaces.

Who it's for

Models at the 2B scale aren't meant to match large models on general ability, and Liquid itself acknowledges that larger models still win on agentic coding and knowledge tasks. They make more sense where data can't leave the device, for batch jobs that need huge numbers of small concurrent agents (Liquid demonstrated more than 300,000 agents working through the Enron email archive in parallel), and for teams willing to spend one single-GPU fine-tuning run turning a small model into an expert on a narrow task. Most benchmark results are self-reported by Liquid, so it's worth running your own tasks before committing.