LFM3-1.7B

LFM3-1.7B is Liquid AI's open-weight 1.73B text model for local agents and specialized applications, the text post-trained checkpoint of the LFM3 family (Oct 27, 2026). New tri-hybrid architecture: 24 layers (13 gated short convolutions, 5 sliding-window attention with 256-token window, 6 global attention), head-wise gated attention, partial RoPE, sandwich norm. Pre-trained on 22T tokens; 256K context; 128K vocabulary; 16 languages. Post-trained (SFT, mixed RL, multi-domain on-policy distillation, agentic RL) for harnesses like Hermes Agent, Pi, and OpenClaw; reasoning on by default; supports deferred tools / tool search. Fits in ~1 GB at 4-bit. Vendor says it is not recommended for agentic coding or complex multi-step reasoning. Hugging Face: LiquidAI/LFM3-1.7B (+ Base, FP8, NVFP4, GGUF, MLX, ONNX). License: LFM Open License v1.0. Catalog benches are self-reported from the vendor table. No catalog id (recorded here only): LiveClaw 45.60, Claw-Eval pass^3 51.40, BrowseComp-Plus (OpenClaw subset) 37.00, EnterpriseOps-Gym 28.20, ToolSandbox 80.36, IFStruct 93.72, LiveCodeBench Pro 25Q2 31.74 (pass rate; catalog id is Elo), AA-Omniscience accuracy 7.07 and non-hallucination 92.93. Launch partners include OpenRouter and Vercel, but no gateway id or price was listed at time of adding, so no provider row. Sources: https://www.liquid.ai/blog/lfm3 · https://huggingface.co/LiquidAI/LFM3-1.7B.