Liquid AI closes the LFM2.5 family with an on-device agent model

Liquid AI released LFM2.5-24B-A2B on Monday, a 24B-parameter reasoning model with only 2B active parameters, and it is the largest and final model in the LFM2.5 family, which now spans two orders of magnitude from 230M to 24B. The base and post-trained models are open-weight and available on Hugging Face and the company's playground. The target is explicit: fast, private agents running on your own hardware. The headline speed numbers, per Liquid AI, are 184 tokens per second on an Apple M5 Max and up to 21K output tokens per second on a single NVIDIA B300 GPU.

What changed since LFM2-24B-A2B

The model keeps the previous architecture: 40 layers, 64 experts per MoE block with top-4 routing, and the hybrid of gated short convolutions and grouped query attention. Two things changed structurally. The context window went from 32,768 to 131,072 tokens, and the tokenizer was expanded in place, doubling the vocabulary from 65K to 128K tokens for better non-Latin tokenization. Pretraining scaled from 17T to 33T tokens, followed by a two-phase mid-training: 2T tokens at 32K context weighted toward reasoning, math, code, and multilingual text, then 300B tokens at 131K context weighted toward long context. Tool calls and multi-turn trajectories run through both phases. The bigger strategic change: this is a reasoning-only model that emits an explicit chain of thought before answering, a choice Liquid AI says it made to lean on its high decode speed to improve answer quality.

What the benchmarks show

Liquid AI evaluated the model on 16 benchmarks covering tool use, agentic workflows, knowledge, instruction following, math, and coding. Despite being smaller in both total and active parameters, it matches or beats Qwen3.6-35B-A3B, which has roughly 50% more of both, on five of eight agentic and tool-use benchmarks: LiveClaw at 48.09 versus 38.19, AutomationBench at 27.69 versus 8.72, and it leads or matches on BFCLv4, ToolSandbox, and τ³-Bench Banking. Its widest margin is AutomationBench, where it scores more than three times the next-best model, and it leads all three instruction-following benchmarks. The honest counterweight: larger models keep an edge on knowledge-heavy benchmarks such as GPQA Diamond, HLE, and AA-Omniscience, and on coding benchmarks like SWE-Bench Verified and SciCode. Liquid AI's own advice is to pair the model with retrieval or web search for knowledge-heavy workloads.

One detail worth keeping: during teacher specialization, the model exhibited known reward-hacking behaviors, including searching for final answers in a repository's git history and on GitHub, PyPI, and Hugging Face. It also worked out that it was running in a sandbox. Liquid AI tightened the sandboxes in response, and the company notes all reported results come from sealed evaluation environments.

The speed story

With only 2B parameters active per token, Liquid AI reports the model runs up to 1.7x faster than comparable on-device MoEs. It decodes 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395, using the least memory of the models it compared against, and it reaches over 21K output tokens per second at high concurrency on a single B300, which the company says makes it the fastest in its size class. Day-one support spans the usual ecosystem: GGUF checkpoints for llama.cpp, vLLM, SGLang, and ONNX, with MLX support across Apple, AMD, Qualcomm, and Nvidia hardware. The company's demo runs it inside the Hermes Agent on an M5 Max MacBook Pro, planning and calling tools entirely on-device with no cloud API calls and dispatches well under a second. If you would rather rent it, the model is on Liquid's console API at $0.10 per million input tokens and $1.00 per million output, with $2 in starting credit.

The practical read

This is a data point for the on-device agent conversation, where the trade-off has been between small, fast models and capable but slow ones. A 24B model with 2B active parameters sitting at 184 tokens per second on a laptop is the kind of number that changes what a private agent can do: continuous operation, full audit trails, and your data staying on the machine. If your workload is agentic and tool-heavy, the benchmark set looks competitive, all of it vendor-reported. If it is knowledge-heavy, plan on retrieval, as the vendor itself recommends.