LFM3-1.7B: a 1.1 GB agent model on a Mac

Liquid AI published LFM3-1.7B on October 27. Six 2B-class models branch from one backbone, and this one handles text: text goes in, answers and tool calls¹ come out.

The way it saves size is the story. Of 24 layers, 13 are short convolutions that read only the last few tokens², 5 are sliding-window attention³ with a 256-token window, and only 6 are global attention, so long context costs far less.

Pre-training used 22T tokens, and the final 3.5T was a mid-training stage weighted toward agentic traces and tool calls, which is why a 2B model scores this high on tool use.

Start with the file you download: at 4 bits it is a little over 1 GB, so a phone or a Raspberry Pi 5 runs it.

The window opens to 256K, and filling 128K costs 1.6 GB of KV cache⁴ where Qwen3.5-4B needs 4.3 GB and MiniCPM5-2B 5.6 GB, so the gap shows in long-context work.

Add 1.1 GB of weights to 1.6 GB of cache and you stay under 3 GB, which leaves room on an 8 GB Mac to keep the model loaded while you work.

Pull the GGUF⁵ and open it with llama-server to get an OpenAI-style API⁶. Set the window with --ctx-size and put the whole model on the GPU⁷ on a Mac with --n-gpu-layers 99.

hf download LiquidAI/LFM3-1.7B-GGUF LFM3-1.7B-Q4_K_M.gguf --local-dir ./lfm3

llama-server -m ./lfm3/LFM3-1.7B-Q4_K_M.gguf \
  --ctx-size 32768 --n-gpu-layers 99 --jinja \
  --temp 0.1 --top-k 50 --repeat-penalty 1.1 \
  --port 8080

Sampling is temperature 0.1, top_k 50 and repetition penalty 1.1 for chat and tool use, and temperature 1.0, top_p 0.95, top_k 20 for maths and science.

Tool calling under llama.cpp needs --jinja, because the chat template shipped with the model is what injects the schemas and lets the response be read back as a call. If memory runs short, lower --n-gpu-layers and offload⁸ part of the model at the cost of slower decoding⁹, so decide first whether to cut context or layers.

Pass function schemas through tools and the template emits XML-style calls. Append the result as a tool message and keep sending the schemas on later requests.

With thousands of tools, defer_loading keeps their definitions out of the prompt¹⁰ until a search tool returns the names. Even a 256K window fills fast if every schema rides along.

Post-training had harnesses such as Hermes Agent, Pi and OpenClaw in mind, and matching the tool-result shape your harness expects is what keeps the loop stable.

MacPaw compared the two on 65 everyday tasks inside real macOS environments: LFM3-1.7B reached 64 percent task success against 48 percent for Qwen3.5-2B, used 58 percent fewer tokens per task, and got stuck in under 1 percent of attempts where the comparison model hit 15 percent.

On-device speed came from llama.cpp at Q4_0 with a 32K window: 1.8 to 2.8 times faster decoding than the next model, with the smallest memory footprint, across two phones, a Raspberry Pi 5, an M4 Max and Ryzen AI.

These are vendor and partner measurements. Matching them locally means matching the quantization¹¹ and the window, and a different harness moves the success rate.

It is not recommended for agentic coding or complex multi-step reasoning; the model card says so, and SWE-bench Verified lands at 43.40 against 51.20 for Qwen3.5-9B.

Knowledge is narrow too: Humanity's Last Exam scores 7.17 without tools and 13.90 with search, so fill knowledge by retrieval and let the model judge and call.

The licence is the LFM Open License v1.0, which is not an open-source licence, so read the commercial terms yourself.