On October 27, 2026, Liquid AI introduced LFM3, a new generation of open-weight Liquid Foundation Models built for specialized intelligence on the edge, according to the company's blog post. The first release ships six models in the roughly 2B-parameter range, all branched from LFM3-1.7B-Base and covering text, vision, audio, omni, and structured decisions. Weights are available today on Hugging Face, the Liquid Playground, and the Liquid AI API.
What launched
LFM3 targets fine-tuning and deployment on phones, laptops, Raspberry Pi boards, and NPUs—not only datacenter GPUs. The shared tri-hybrid backbone mixes 13 gated short-convolution layers, 5 sliding-window attention blocks with a 256-token window, and 6 global attention layers across 24 total. Liquid pre-trained the base on 22T tokens, pushed text context to 256K, and expanded hardware-in-the-loop search to 14 device and backend pairings—including Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI, Arm and x86 CPUs, GPUs, and Qualcomm Hexagon NPU.
Liquid says LFM3-1.7B beats its LFM2.5-2.6B with 35% fewer parameters, fits in about 1 GB at 4-bit precision, and that every model fine-tunes on a single GPU.
The six models
LFM3-1.7B is the text model for local agents (256K context), post-trained for harnesses such as Hermes Agent, Pi, and OpenClaw. LFM3-VL-2B handles text, images, and video with step-by-step visual reasoning. LFM3-Audio-2B accepts text or audio and replies in text, speech, or both across 18 languages. LFM3-Omni-2B (about 2.2B parameters) is Liquid's first omni checkpoint—spoken visual QA, audio-visual video, and multimodal tool calls. d1-2B and d1-Omni-2B are decision models on the vision and omni bases: they score batches of questions (yes/no, choices, boxes, extracted fields, and more) in one forward pass instead of writing long answers, with d1-Omni adding voice inputs.
Why small and on-device matters
Liquid's pitch is "specialized intelligence"—choose the modality, adapt on your data, and run where the application lives. Partners including AMD, Arm, Qualcomm, and Mercedes-Benz highlight AI PCs, edge vision-language stacks, and in-car assistant experiments. Operationally, Liquid reports a full 128K-token KV cache at about 1.6 GB for LFM3-1.7B versus 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B in llama.cpp Q4_0 tests. MacPaw cites 64% versus 48% task success against Qwen3.5-2B on 65 macOS agent tasks, with 58% fewer tokens and under 1% stuck runs versus 15%.
Benchmarks—with caveats
Agent tables favor the small text model on browsers and ops work: WebArena-Verified 38.42 versus 28.94 for Qwen3.5-9B, and AutomationBench (public) 41.29 versus 15.10. Liquid also highlights strong instruction-following and tool-calling scores on BFCL and ToolSandbox relative to much larger checkpoints.
The blog states plainly that larger models still win on agentic coding and knowledge, where size matters most—SWE-bench Verified (OpenCode) is 43.40 for LFM3-1.7B versus 51.20 for Qwen3.5-9B. On vision, LFM3-VL-2B averages 71.4 across 62 benchmarks, nearly matching Gemma 4 12B at 71.7. Decision scores such as Decision Bench at 85.2 for d1-2B versus 83.6 for d1-3B come from Liquid's own enterprise-oriented suites as well as public tests.
Availability
Weights and API access are live. Day-0 partners include Ambarella, AMD, Arm, Intel (OpenVINO), NVIDIA, OpenRouter, Qualcomm (AI Hub and GenieX), and Vercel, with llama.cpp, MLX, vLLM, and SGLang called out across Apple, AMD, NVIDIA, and Qualcomm hardware. Mercedes-Benz and Shopify helped pre-release testing; builders can download from Hugging Face, try the Playground, or route traffic through Liquid's Real-time Inference Pipeline.
What it means
LFM3 packages multimodal edge AI as a menu of open checkpoints on one backbone—agents, camera apps, voice UX, omni assistants, and fast decision heads—without assuming a single cloud generalist. Benchmark caveats included: do not expect 1.7B text weights to top 9B coders on SWE-bench, but they may fit privacy, offline, or per-device budgets frontier APIs cannot. For teams already shipping on-device, day-0 silicon support and single-GPU fine-tunes make this a low-friction experiment; production proof will come from partners like Shopify and Mercedes-Benz, not leaderboard rows alone.