Liquid AI Ships LFM2.5-24B-A2B at 184 tok/s on M5 Max

Liquid AI's new reasoning model, LFM2.5-24B-A2B, has 24 billion total parameters but activates only 2.45 billion of them per token. The model card on Hugging Face frames it as "a 24B-parameter reasoning model with only 2B active parameters, built for fast, private agents on-device."

The architecture pairs 30 double-gated short-convolution blocks with 10 GQA layers, running 64 experts per MoE block with top-4 routing. Context stretches to 131,072 tokens. The training budget was 36 trillion tokens: roughly 33T of pre-training, 2T of mid-training at 32K context, and another 300B at the full 131K window.

On an Apple M5 Max, Liquid clocks 184 output tokens per second, which it calls "1.7× faster than comparable MoEs" on-device. An AMD Ryzen AI Max+ 395 runs it at 114 tok/s. A single NVIDIA B300 can push up to 21,000 output tokens per second at high concurrency, tested with SGLang 0.5.22 at 1,024 input tokens and up to 256 output tokens in BF16.

The company compares the model directly against Qwen3.6-35B-A3B, which has roughly 50% more total and active parameters. LFM2.5 leads on tool-use and instruction-following: LiveClaw 48.09 vs 38.19, AutomationBench 27.69 vs 8.72, ToolSandbox 80.71 vs 75.02, IFEval 96.75 vs 93.06, IFBench 74.69 vs 64.21. Qwen3.6 wins on reasoning: MATH-500 96.73 vs 96.04, GPQA Diamond 83.69 vs 74.42, and WebArena Verified 37.62 vs 26.75.

Weights ship in native, FP8, GGUF, ONNX, and MLX-4bit variants under the lfm1.0 license. Liquid writes that "with only 2B active parameters per token, each reasoning token is cheap." The card does not disclose any hosted API price.

The drop lands in a crowded run of open-weights MoE releases; our open-source tracker counts 341 stories in the last 90 days.