Liquid AI Releases LFM2.5-24B-A2B: A 24B-Parameter Reasoning Model with 2B Active Parameters for On-Device Agents

Liquid AI has released LFM2.5-24B-A2B, a 24-billion parameter reasoning model designed for fast, private on-device agents. Built on a Mixture-of-Experts (MoE) architecture with only 2B active parameters per token, the model balances high reasoning capability with low computational cost, enabling continuous agent operation on local hardware.

Architecture and Training

The model builds on the LFM2-24B-A2B architecture, featuring 40 layers and 64 experts per MoE block with top-4 routing. Key upgrades include:

• Context Window: Expanded from 32K to 131K tokens.
• Tokenizer: Vocabulary doubled from 65K to 128K tokens via in-place expansion for better non-Latin support.
• Training Data: Pre-training scaled to ~33T tokens, followed by mid-training on 2T tokens at 32K context and 300B tokens at 131K context.
• Post-Training: A five-stage process including supervised fine-tuning, teacher specialization, multi-domain on-policy distillation (MOPD), and agentic reinforcement learning within harnesses like OpenClaw and Hermes Agent.

Benchmark Performance

LFM2.5-24B-A2B competes with models having ~50% more total and active parameters, such as Qwen3.6-35B-A3B. It leads or matches competitors in tool use and instruction following:

• LiveClaw: Scored 48.09, compared to 38.19 for Qwen3.6-35B-A3B.
• AutomationBench: Scored 27.69, more than three times the next-best model.
• BFCLv4: Scored 65.01, slightly ahead of Qwen3.6-35B-A3B (65.00).
• Instruction Following: Leads on IFEval (96.75), IFBench (74.69), and Multi-IF (84.38).

Larger models retain advantages in knowledge-heavy tasks (GPQA Diamond, HLE) and coding (SWE-Bench Verified), suggesting LFM2.5-24B-A2B is best paired with retrieval or web search for those workloads.

Inference Speed

The model offers significant speed advantages due to its low active parameter count:

• Edge Devices: Decodes at 184 tokens/s on an Apple M5 Max and 114 tokens/s on a Ryzen AI Max+ 395, up to 1.7x faster than comparable MoEs.
• GPU: Reaches over 21K output tokens/s at high concurrency on a single NVIDIA B300 GPU.

Availability

Both LFM2.5-24B-A2B and the base version are open-weight and available on Hugging Face. The model supports day-one integration with llama.cpp, MLX, vLLM, SGLang, and ONNX, and requires transformers>=5.0.0.