Run fast, private agents on-device with LFM2.5-24B-A2B

Liquid AI has put out LFM2.5-24B-A2B, a reasoning model aimed at agents that can run locally instead of in the cloud. The headline number is a bit of a trick: it has 24B parameters overall, but only 2B are active for each token. That keeps reasoning cheaper, which is the whole point if you want something that can stay on your own machine and keep working.

The model is not just a smaller fork of something bigger. It keeps the same basic setup as LFM2-24B-A2B, with 40 layers, 64 experts per MoE block, top-4 routing, and the same mix of gated short convolutions and GQA. The context window also jumps from 32K to 131K tokens, and the tokenizer expands in place from 65K to 128K tokens so non-Latin text is handled better. Unlike the earlier model, this one is reasoning-only and writes out an explicit chain of thought before answering.

Liquid AI says it scaled pre-training from about 17T tokens to about 33T, then added 2T tokens at 32K context and 300B at 131K context during mid-training. Those later stages lean into reasoning, math, code, instruction following, multilingual text, and long-context data. Tool calls and multi-turn trajectories show up throughout both phases, and the post-training pipeline ends with supervised fine-tuning, teacher specialization, two rounds of multi-domain on-policy distillation, and agentic reinforcement learning inside harnesses like OpenClaw, Hermes Agent, Pi, and OpenCode.

The pitch is supported by benchmark results that favor tool use and instruction following. On LiveClaw, LFM2.5-24B-A2B scores 48.09, ahead of Qwen3.6-35B-A3B at 38.19. It also leads or matches the comparison set on AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking, and it tops all three instruction-following benchmarks in the release. Larger models still keep the edge on knowledge-heavy and coding tests such as GPQA Diamond, HLE, AA-Omniscience, SWE-Bench Verified, and SciCode, so Liquid AI is explicit about pairing this model with retrieval or web search for those jobs.

Speed is the other half of the story. Liquid AI says LFM2.5-24B-A2B decodes 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395, with up to 21K output tokens per second on a single NVIDIA B300 at high concurrency. The model ships with day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX, so this is clearly meant to be used, not just admired in a blog post.