LFM2.5-24B-A2B: a local agent runs on a laptop CPU at 184 tokens per second

Of its 24 billion parameters, 2.45 billion are active per token, so the model generates 184 tokens per second on a CPU.

A 24-billion-parameter model answers straight from a laptop CPU at 184 tokens per second.

On Oct 13, Liquid AI released LFM2.5-24B-A2B on Hugging Face. Of its 24 billion parameters, 2.45 billion wake up for each token, so the giant computes at the speed of a mid-size model. The vendor's benchmark on an Apple M5 Max gives 184 tokens per second; on an AMD Ryzen AI Max+ 395 it reaches 114.

For a practitioner this means one thing: the agent runs on your own hardware, with no cloud and no token bill. The model knows 16 languages.

How to run it. One command in llama.cpp: `llama-cli -hf LiquidAI/LFM2.5-24B-A2B-GGUF --temp 0.1 --top-k 50 --repeat-penalty 1.05`. In LM Studio, the model card installs it in one click. The recommended Q4_K_M quant weighs 14.5 GB, and the lightest, Q4_0, takes 13.6 GB. An MLX-4bit variant is available for Apple Silicon.

How to connect an agent. The model answers at an OpenAI-compatible endpoint, so a local agent is set up with four lines of config. In Hermes Agent these are `model.provider custom`, `model.base_url http://localhost:8080/v1`, `model.default LFM2.5-24B-A2B` and `model.context_length 131072`. Liquid AI published the same guides for OpenClaw and Pi.

The previous LFM2-24B-A2B held 32,768 tokens of context. The new version takes 131,072 and was trained on 33 trillion tokens instead of 17. The style of answering changed too: the model always thinks out loud, and the chat template turns on the chain of thought by itself.

By the vendor's measurements, the model scores 65.01 on BFCLv4 and 80.71 on ToolSandbox, on par with Qwen3.6-35B-A3B with half as many active parameters. On code the gap shows. SWE-Bench Verified via OpenCode gives 51.00% versus 66.20% for Qwen, and for knowledge tasks Liquid AI advises falling back on document or web search.

Five runtimes supported the model on launch day: llama.cpp, MLX, vLLM, SGLang and ONNX.

LFM2.5-24B-A2B: a local agent runs on a laptop CPU at 184 tokens per second

Of its 24 billion parameters, 2.45 billion are active per token, so the model generates 184 tokens per second on a CPU.

A 24-billion-parameter model answers straight from a laptop CPU at 184 tokens per second.

For a practitioner this means one thing: the agent runs on your own hardware, with no cloud and no token bill. The model knows 16 languages.

Five runtimes supported the model on launch day: llama.cpp, MLX, vLLM, SGLang and ONNX.