Liquid AI's LFM2.5 Fits a Reasoning Agent in 2B Parameters
Liquid AI's new open-weight model beats larger MoEs on agentic tasks and runs at 184 tokens/s on a laptop chip.
By The ai-compute Desk
Liquid AI has released LFM2.5-24B-A2B, an open-weight mixture-of-experts model that activates only 2B parameters per token yet scores 48.09 on LiveClaw—a tool-use benchmark where the larger Qwen3.6-35B-A3B, carrying roughly 50% more total and active parameters, manages 38.19. [1] The model is available now on Hugging Face.
The architecture inherits 40 layers and 64 experts per MoE block with top-4 routing from its predecessor, LFM2-24B-A2B, but two changes define the new release. [1] First, the context window expands from 32K to 131K tokens. Second, an in-place tokenizer expansion doubles the vocabulary from 65K to 128K tokens for better non-Latin tokenization. [1]
That tokenizer upgrade matters beyond Latin-script users. Hindi and Vietnamese now require roughly 2.4× and 2.6× fewer tokens respectively, and Thai up to 4.0× fewer, which Liquid estimates makes per-character decoding 2.2 to 3.7× faster for those languages on-device. [2]
Pre-training for LFM2.5-24B-A2B was scaled from roughly 17T to 33T tokens. Mid-training added 2T tokens at 32K context weighted toward reasoning, math, code, instruction following, and multilingual text, then 300B tokens at 131K context weighted toward long-context data. [1]
On the benchmark sweep across 16 tests covering tool use and agentic workflows, knowledge, instruction following, math, and coding, LFM2.5-24B-A2B leads or matches competitors on five of the eight tool-use and agentic benchmarks—AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking—and on AutomationBench it scores more than three times the next-best model. [1] It also leads on all three instruction-following benchmarks. [1]
The results sharpen a tension that has been building across the edge-AI field: raw parameter count is an increasingly poor proxy for useful capability. What matters on a phone or a laptop is the ratio of useful work to memory bandwidth consumed—and that ratio is where compact, sparsely-activated architectures are starting to pull ahead of their denser rivals.
Speed numbers back that claim. On an Apple M5 Max, LFM2.5-24B-A2B decodes at 184 tokens per second—up to 1.7× faster than the next-best model in the comparison group, while using the least memory. [1] On a Ryzen AI Max+ 395 it reaches 114 tokens per second, and at high concurrency on a single NVIDIA B300 GPU it exceeds 21K output tokens per second. [1] Day-one inference support covers llama.cpp, MLX, vLLM, SGLang, and ONNX. [1]
[2]
Both decisions—where to stop expanding the vocabulary and which benchmarks to optimize for—reflect a disciplined set of trade-offs rather than a grab for headline numbers. That kind of restraint is what separates models designed for real deployment from those designed for leaderboard position.
LFM2.5-24B-A2B and its base variant are open-weight and available on Hugging Face today, alongside fine-tuning tutorials and deployment guides for every major inference backend. [1] The question is no longer whether a reasoning agent can fit on your device—it's what you intend to do with one that already can.
It is the smallest model in the group, yet it is competitive with MoEs with ~50% more total and active parameters.Liquid AI