Liquid AI released LFM2.5-24B-A2B on October 13, an open-weight reasoning model with 24 billion parameters in total and 2 billion active for each token, built to run agents on local hardware. In its release post, Liquid reports 184 tokens per second of decoding on an Apple M5 Max and 114 on a Ryzen AI Max+ 395, and over 21K output tokens per second at high concurrency on a single NVIDIA B300. The post-trained model and the base model are both on Hugging Face, with day-one support in llama.cpp, MLX, vLLM, SGLang and ONNX.
Only 2B parameters work per token, but all 24B still have to sit in memory
The model keeps the architecture of LFM2-24B-A2B: 40 layers, 64 experts per mixture-of-experts block with the top four selected for each token, and a hybrid of gated short convolutions and grouped-query attention. What changes is the context window, which grows from 32K to 131K tokens, the vocabulary, which doubles from 65K to 128K tokens, and the pre-training run, which rises from about 17 trillion to about 33 trillion tokens. Unlike its predecessor, it always writes out a chain of thought before answering.
The small active count is what makes each token cheap, but every expert has to be loaded. At 2 bytes per parameter, the bfloat16 precision used in Liquid’s own loading example, the weights alone would likely take about 48GB, before any context cache. A 24GB graphics card would likely need a lower-precision build, and the post does not say which quantization produced the M5 Max and Ryzen figures.
It leads on tool use and instruction following, while the larger Qwen3.6-35B-A3B stays ahead on coding and knowledge
Liquid tested it on 16 benchmarks against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning, all run on vLLM with each benchmark’s recommended settings. It scores 48.09 on LiveClaw against 38.19 for Qwen3.6-35B-A3B, 80.71 on ToolSandbox against 75.02, and 27.69 on AutomationBench against 8.72. On BFCLv4 the two are level at 65.01 and 65.00. It also leads all three instruction-following tests, with 74.69 on IFBench against 64.21 for Qwen3.6-35B-A3B.
Qwen3.6-35B-A3B keeps the lead on coding and knowledge. It scores 66.20 on SWE-Bench Verified run through OpenCode, against 51.00 for Liquid’s model, and 83.69 on GPQA Diamond, against 74.42. Liquid’s own advice is to pair the model with retrieval or web search for knowledge-heavy work.
Every score comes from Liquid’s own runs. The text of the post gives no memory figure, no quantized file sizes and no context length for the speed numbers, so how fast it decodes deep into a long agent session is not yet known.