LFM3-1.7B

// Overview

LFM3-1.7B is the text model of LFM3, the third generation of Liquid Foundation Models, announced on the Liquid AI blog on October 27, 2026. It has 1.73B parameters, a 256K-token context window (input plus output) and a 128,000-token vocabulary, and it supports 16 languages including English, Chinese, Arabic, Hindi, Japanese and Russian. Liquid describes it as a model for local agents, tool calling, structured outputs and domain fine-tuning: at 4-bit precision it takes about 1 GB of memory. Reasoning is on by default and can be switched off with enable_thinking=False.

LFM3 introduces a tri-hybrid architecture. Of its 24 layers, 13 are gated short convolutions (kernel size 7, up from 3 in LFM2), 5 are sliding-window attention layers with a 256-token window and 6 are global attention layers. Pre-training covers 22T tokens; the last 3.5T are curated mid-training data with more agentic traces and tool calls, which also extends the context to 256K. Post-training runs 1.75T tokens of supervised fine-tuning, mixed reinforcement learning across many domains, multi-domain on-policy distillation from six specialist RL teachers (code, math, tool use, long context, agentic work and enterprise workflows), and agentic RL inside real agent harnesses. The same LFM3-1.7B-Base also branches into LFM3-VL-2B (vision), LFM3-Audio-2B, LFM3-Omni-2B and two decision models.

On Liquid's 29-benchmark comparison it averages 49.37, ahead of Qwen3.5 9B (47.24), Qwen3.5 4B (39.27), MiniCPM5 2B (39.19), Gemma 4 E4B (31.74) and LFM2.5-2.6B (31.55). Its leads are in agents and tool use — AutomationBench 41.29 against 15.10 for Qwen3.5 9B, WebArena-Verified 38.42 against 28.94, BFCL V4 71.39 against 62.83 — and in instruction following (IFEval 96.66, IFBench 73.03). Liquid is explicit that larger models still win on knowledge and agentic coding: Qwen3.5 9B scores 80.40 on GPQA Diamond to its 63.37 and 51.20 on SWE-bench Verified (OpenCode) to its 43.40, and the model card advises retrieval or search for knowledge-heavy tasks and does not recommend it for agentic coding.

On Liquid's edge benchmarks (llama.cpp, Q4_0, 32K context) it decodes 237.1 tokens/s on an Apple M5 Max, 129.6 on an AMD Ryzen AI Max+ 395 and 43.6 on an iPhone 17 Pro, using the least memory of the compared models (2.25 GB on the M5 Max); a full 128K-token context needs 1.6 GB of KV cache. On one NVIDIA B300 with SGLang it can produce 3.6 billion tokens per day at 1,024 concurrent requests. Weights ship under the LFM Open License v1.0, with support in Transformers, vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX and ONNX Runtime.