Liquid AI Releases LFM2.5-24B-A2B: A 24B-Parameter MoE Model Optimized for Fast On-Device Agents

Liquid AI has released LFM2.5-24B-A2B, the largest and final model in its LFM2.5 family. This 24B-parameter Mixture-of-Experts (MoE) model is designed for fast, private agent execution on local hardware, activating only 2B parameters per token. It is now available as open-weight on Hugging Face and via Liquid AI's Playground.

Performance and Benchmarks

LFM2.5-24B-A2B demonstrates competitive performance against larger models, specifically outperforming or matching Qwen3.6-35B-A3B (which has ~50% more total and active parameters) on five of eight agentic and tool-use benchmarks.

• AutomationBench: Scored 27.69 vs. 8.72 for Qwen3.6-35B-A3B.
• LiveClaw: Scored 48.09 vs. 38.19 for Qwen3.6-35B-A3B.
• BFCLv4: Scored 65.01, effectively tied with Qwen3.6-35B-A3B (65.00).
• Instruction Following: Leads on IFEval (96.75), IFBench (74.69), and Multi-IF (84.38).

While it excels in tool use and instruction following, larger models retain advantages in knowledge-heavy tasks (GPQA Diamond, HLE) and complex coding benchmarks (SWE-Bench Verified).

Architecture and Training

The model builds on the previous LFM2-24B-A2B with significant architectural and training updates:

• Context Window: Extended from 32,768 to 131,072 tokens.
• Vocabulary: Doubled from 65K to 128K tokens via in-place tokenizer expansion for better non-Latin support.
• Pre-training: Scaled from 17T to 33T tokens.
• Reasoning: Unlike its predecessor, LFM2.5-24B-A2B is a reasoning-only model that generates an explicit chain of thought before answering.
• Post-training: Utilizes a five-stage pipeline including Supervised Fine-Tuning (SFT), Teacher Specialization, Multi-Domain On-Policy Distillation (MOPD), and Agentic Reinforcement Learning (Agentic RL) within harnesses like OpenClaw and Hermes Agent.

Inference Speed and Deployment

Due to its efficient architecture, LFM2.5-24B-A2B offers high throughput on consumer and server hardware:

• Apple M5 Max: Decodes at 184 tokens/s.
• AMD Ryzen AI Max+ 395: Decodes at 114 tokens/s.
• NVIDIA B300 GPU: Reaches over 21K output tokens/s at high concurrency.

The model ships with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. It is priced at $0.10/1M input tokens and $1.00/1M output tokens via the Liquid Console API.