LFM3-1.7B

An independent writeup of the October 27, 2026 LFM3 launch reports that the text model LFM3-1.7B averages 49.37 on Liquid's 29-benchmark table versus 47.24 for Qwen3.5-9B, and decodes 43.6 tokens per second on an iPhone 17 Pro using 1.90 GB of memory, roughly 2.8x faster than the next model Liquid tested. The piece also flags practical constraints: Liquid's model card says not to use LFM3-1.7B for agentic coding or complex multi-step reasoning, weights on Hugging Face are free only for companies under $10M in annual revenue, and no API price for LFM3 was published at launch. All benchmark numbers are vendor-reported.

This technical deep dive verifies the LFM3-1.7B configuration against the released config.json and reference implementation: 24 layers comprising 13 gated short convolutions with kernel size 7, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers, alongside head-wise gated attention, partial RoPE, and sandwich norm. Post-training runs through four stages ending in agentic reinforcement learning inside real harnesses. Reported figures, quoted from Liquid's own published tables, put LFM3-1.7B at 49.37 average against 47.24 for the five-times-larger Qwen3.5-9B, with class-leading agent scores on LiveClaw, AutomationBench, and WebArena-Verified. The model also carries a 256K context window, fits in about 1 GB at 4-bit, and decodes 1.8 to 2.8 times faster than the next-fastest model on Liquid's measured edge devices.

Liquid AI's LFM3 release spans six compact models sharing a 1.7-billion-parameter foundation, with specialized variants near 2B parameters, and the company says enterprises can customize each on a single GPU and run them on devices, private servers, or cloud infrastructure. Company-reported results suggest the smallest text model can beat much larger rivals on selected browser automation and tool-use tasks. The flagship LFM3-1.7B text model documents a 256K-token combined input and output context window, while Liquid's hosted API console lists a 128,000-token limit, a distinction that matters for hosted versus self-managed deployments. Liquid also reports the text model trails Qwen3.5-9B on several coding, knowledge, and long-context evaluations.

All six LFM3 models branch from the same LFM3-1.7B-Base backbone, with LFM3-1.7B handling text, LFM3-VL-2B accepting text, images, and video, LFM3-Audio-2B for text and audio, LFM3-Omni-2B as Liquid's first omni model, and d1-2B and d1-omni-2B returning calibrated decisions instead of generated text. The tri-hybrid design uses convolution kernel size 7, up from 3 in LFM2, plus head-wise gated attention, partial RoPE, and a sandwich norm. Day-0 launch partners include Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, OpenRouter, Qualcomm via Qualcomm AI Hub and GenieX, and Vercel, while Mercedes-Benz and Shopify handled pre-release testing.

Liquid AI announced LFM3, its third generation of Liquid Foundation Models, on October 27, 2026, built as a multimodal family on one shared backbone covering text, vision, audio, omni, and decision models in the 2B parameter range. The backbone was pre-trained on 22T tokens, and base plus post-trained weights are open and available on Hugging Face, the Liquid Playground, and the Liquid API. The architecture comes from Liquid's automated hardware-in-the-loop search, expanded for LFM3 to 14 device and backend combinations including Samsung Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI, and the Qualcomm Hexagon NPU. Gated short convolutions still form over half the layers, with sliding-window attention giving the best quality-speed tradeoff, so LFM3-1.7B is about 25% faster at 8K context.

Liquid AI documents LFM3-1.7B as a 1.73B-parameter dense text model built on the LFM3 tri-hybrid architecture. Its 24 layers combine gated short convolution, sliding-window attention, and global attention, while a 256K-token context window supports long tool traces and multi-step workflows. The model targets local agents, tool calling, structured outputs, and domain-specific fine-tuning, and is post-trained for harnesses including Hermes Agent, Pi, and OpenClaw. At 4-bit precision it requires about 1 GB of memory, and official distributions include GGUF, MLX, ONNX, FP8, and NVFP4 formats under LFM Open License v1.0.

The official LFM3-1.7B model card describes a post-trained, text-only model with 1.73B parameters across 24 layers: 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers. It supports a 256K-token context window, a 128,000-token vocabulary, and 16 languages, and ships under the LFM Open License v1.0. The model targets local agents, tool calling, structured outputs, and domain fine-tuning, post-trained for harnesses such as Hermes Agent, Pi, and OpenClaw. At 4-bit precision it needs about 1 GB of memory, and checkpoints cover BF16, FP8, NVFP4, GGUF, MLX, and ONNX. The card explicitly advises against agentic coding or complex multi-step reasoning.