LiquidAI released LFM2.5-24B-A2B on October 13, 2026, an open-weight reasoning model carrying 24 billion total parameters but activating only 2.45 billion per token, targeting on-device agentic workloads1,3. Both the base (LFM2.5-24B-A2B-Base) and post-trained checkpoints are available on Hugging Face Inc. and LiquidAI's Playground, with day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX2.
The model is the largest and final entry in the LFM2.5 family, which now spans from LFM2.5-230M to LFM2.5-24B-A2B.
Architecture and what changed
LFM2.5-24B-A2B retains the architecture of its predecessor, LFM2-24B-A2B: 40 layers composed of 30 double-gated short convolution blocks and 10 grouped-query attention blocks, with 64 experts per mixture-of-experts block and top-4 routing. Two structural changes separate it from LFM2-24B-A2B: the context window grew from 32,768 to 131,072 tokens, and an in-place tokenizer expansion doubled the vocabulary from 65,000 to 128,000 tokens to improve non-Latin tokenization. The training budget is listed at 36 trillion tokens. Unlike its predecessor, LFM2.5-24B-A2B is a reasoning-only model that produces an explicit chain of thought before its final answer, with a ChatML-style chat template inserting a think tag at the start of each assistant response.
The model supports 16 languages, including English, Arabic, Chinese, French, German, Japanese, Korean, Spanish, Hindi, and Polish.
Benchmark and speed claims
LiquidAI positions LFM2.5-24B-A2B against Qwen3.6 35B A3B, a model with roughly 50% more total and active parameters. On five of eight agentic and tool-use benchmarks, LiquidAI reports the model is competitive with Qwen3.6 35B A3B, scoring 48.09 on LiveClaw versus 38.19 and 27.69 on AutomationBench versus 8.72.
Because only 2B parameters fire per token, LiquidAI claims inference runs up to 1.7× faster than comparable MoEs on-device. Reported throughput figures include 184 tokens per second on an Apple M5 Max and up to 21,000 output tokens per second on a single NVIDIA B300 GPU.
LiquidAI lists API pricing at $0.10 per million input tokens, $0.05 per million cached input tokens, and $1.00 per million output tokens.
ANALYSIS The 2.45B active-parameter footprint at 24B total is an aggressive sparsity ratio, and the throughput numbers on consumer hardware position the model squarely at practitioners who want to run agentic loops locally rather than through a cloud endpoint. The pricing, meanwhile, undercuts most hosted reasoning-model tiers by a wide margin.