LFM3-1.7B：新機能、必要なハードウェア、無料での動かし方

LFM3-1.7B packs a 256K token window into about 1 GB of RAM. It’s built for local agents and can run on a phone or laptop.

LFM3-1.7B is a text-only language model from LiquidAI. The architecture mixes gated short convolutions, sliding-window attention and global attention across 24 layers. The model is released under the LFM Open License v1.0.

According to LiquidAI, the model is post-trained for agentic use cases like Hermes Agent, Pi and OpenClaw, and it comes with a 256 K token context window. At 4-bit precision it needs only about 1 GB of memory, so it fits on any phone or laptop. The design encourages quick domain-specific fine-tuning on a single GPU, making it easy to adapt to new formats or tools.

The model card lists several benchmark results. In the LiveClaw agent test LFM3-1.7B scores 45.60, beating larger models like Qwen3.5-9B. For tool-use tasks it reaches 71.39 on BFCL V4. In coding, LiveCodeBench v6 shows a 73.46 score, and on instruction following IFEval it hits 96.66. These numbers indicate strong performance for a 1.7 B parameter model.

LFM3-1.7B shines in agentic applications, tool calling and structured output tasks, and it runs faster than comparable models on edge devices. However, the model card warns that it is not recommended for agentic coding or complex multi-step reasoning because of its small size.

If you run AI at home, need a model that fits on a phone or laptop, or want to fine-tune quickly for a specific domain, LFM3-1.7B is a good fit. It works on modest GPUs like an RTX 4060 and can be run locally with GGUF files in llama.cpp, Ollama or LM Studio.

自宅で動かすには何が必要でしょうか。4ビットでは約4ギガバイトのメモリを使います。完全に収まる最小の家庭用マシンはRTX 4060です。RTX 3060 12 GBとRAM 32ギガバイトでは、毎秒約199トークンで書きます。RTX 4090とRAM 64ギガバイトでは、毎秒約410トークンで書きます。

There is no free API for it yet. The weights are open: you can download them from Hugging Face, under the license written in its model card. Arena has not ranked it yet: we add the score as soon as the votes arrive.

Compare LFM3 1.7B with every free model, and check what your own GPU can run, on FreeLLMArena.