1.73B text/base; 1.92B vision and d1; 2B audio; 2.17B omni and d1-Omni
256K text; 128K vision and decision models; 32K audio and omni
LFM Open License v1.0 (commercial use permitted below US$10 million annual revenue; separate license required at or above the threshold)
Same-day open-weight family of compact text, vision, audio, omni-modal, and non-generative decision models built on Liquid AI's new tri-hybrid backbone. The post-trained LFM3-1.7B text checkpoint contains exactly 1,728,801,536 BF16 parameters in a single approximately 3.46 GB safetensors file and uses 24 layers comprising 13 gated short-convolution blocks, five 256-token sliding-attention layers, and six global-attention layers. Its 2,048-dimensional backbone has 7,680-dimensional feed-forward layers, 16 query heads, four key/value heads, 128-dimensional attention heads, sandwich normalization, tied embeddings, and a 128,000-token vocabulary. Full-attention layers use 50% partial rotary embeddings with a 1,000,000 RoPE base, while sliding layers use full rotary embeddings and a 10,000 base. The text model supports 262,144 input-plus-output tokens, thinking enabled by default, optional preservation of earlier reasoning, XML tool calls, deferred tool definitions and tool search, structured output, and sixteen written languages. Liquid positions it for local agents and domain specialization rather than knowledge-heavy tasks, agentic coding, or highly complex multistep reasoning. The 1.92B LFM3-VL-2B combines the 1.73B language backbone with a 164M in-house vision encoder and 25M projector, accepts native-resolution images and two-frame-per-second video, and supports 128K tokens. LFM3-Audio-2B accepts text or 16 kHz audio and generates text, 24 kHz speech, or interleaved text and speech across eighteen speech locales with a 32K context. The 2.17B LFM3-Omni-2B accepts text, images, video with or without its soundtrack, and audio, while producing text, speech, or both within 32K tokens. Companion d1-2B and d1-Omni-2B models answer named classification, extraction, routing, scoring, and inspection questions in one forward pass, returning calibrated typed values and probabilities rather than generating tokens; d1-Omni additionally accepts spoken questions and audio clips. Pretraining uses 22T tokens: approximately 15T web and synthetic tokens, 3T code, mathematics, and science tokens, 3T curated agentic mid-training tokens at 32K, 500B tokens extending context to 128K, and 50B extending it to 256K. The last 3.5T tokens emphasize higher-quality agent traces and tool calls. Post-training comprises 1.75T supervised-finetuning tokens, mixed reinforcement learning with a concision reward, Multi-Domain On-Policy Distillation from six code, mathematics, tool-use, long-context, agentic, and enterprise specialists, and final agentic reinforcement learning inside real harnesses including Hermes Agent, Pi, and OpenClaw. Liquid reports a 49.37 unweighted average across its text benchmark suite, versus 47.24 for Qwen3.5-9B, 39.27 for Qwen3.5-4B, 39.19 for MiniCPM5-2B, and 31.55 for LFM2.5-2.6B. Individual reported results include 45.60 on LiveClaw, 68.64 on Claw-Eval General, 41.29 on AutomationBench, 71.39 on BFCL V4, 80.36 on ToolSandbox, 68.01 on tau3-bench, 43.40 on SWE-bench Verified, 73.46 on LiveCodeBench v6, 96.66 on IFEval, and 65.56 on RULER 128K. These are publisher-run evaluations served in BF16 at each model's full context length. At 32K using Q4_0 GGUF, Liquid reports at least 2.5 times the prefill throughput and 1.8 times the decode throughput of every compared model across AMD Ryzen AI Max+ 395, Apple M5 Max, and iPhone 17 Pro tests; at 1,024 concurrent requests, it reports 3.6 billion generated tokens per day on one NVIDIA B300. Official deployment formats across the family include BF16, FP8, NVFP4, GGUF, MLX, and ONNX.