2026-10-14 13:49

Introduction: What is LFM2.5-24B-A2B? LFM2.5-24B-A2B is an open-weight inference model launched by Liquid AI, belonging to the flagship LFM2.5 series.

What is LFM2.5-24B-A2B?

LFM2.5-24B-A2B is an open-weight inference model introduced by Liquid AI, serving as a flagship in the LFM2.5 series. With a total of 24B parameters and only 2.45B activated per token, it features a 40-layer MoE architecture, supports up to 131K context length, and handles 16 languages. The model generates a chain-of-thought prior to answering, specifically optimized for edge-side Agents. It outperforms Qwen3.6-35B-A3B—whose parameter count is 50% larger—in LiveClaw scoring (48.09 vs. 38.19). On M5 Max, it achieves decoding speeds of 184 tok/s, 1.7× faster than the fastest comparable MoE model. Ideal for local, private deployment scenarios involving inference and tool calling, it is recommended to pair with retrieval for knowledge-intensive tasks.

Key Features of LFM2.5-24B-A2B

• Edge-Side Agent Inference: Designed for fast, privacy-preserving intelligent agents running locally on personal devices.
• Pure Inference Mode: Automatically generates explicit chain-of-thought before responding, achieving high answer quality at low activation cost.
• Function Calling: Native support for a four-step tool calling pipeline, directly integrable into frameworks such as OpenClaw, Hermes Agent, Pi, and OpenCode.
• Multi-Format Deployment: Available in native, FP8, GGUF, ONNX, and MLX-4bit versions, supporting GPU, CPU, Apple Silicon, and edge devices.
• Long Context Handling: Supports 131,072 tokens of context, enabling processing of long documents and multi-turn agent trajectories.
• Multilingual Capability: Covers 16 languages including English, Chinese, Japanese, Korean, Arabic, with a 128K vocabulary optimized for non-Latin script tokenization.
• High-Speed Inference: Achieves 184 tok/s on Apple M5 Max—1.7× faster than the fastest comparable MoE model—and over 21K tok/s on single B300 GPU.

Technical Principles of LFM2.5-24B-A2B

• MoE Sparse Activation Architecture: While the model has 24B total parameters, each MoE block contains 64 experts. During inference, only 4 experts are activated via top-4 routing per token, resulting in actual computation of just 2.45B parameters per token. This "high-capacity, low-compute" design allows the model to store all knowledge while activating only a small fraction during forward passes, drastically reducing inference cost compared to models with similar total parameter counts.
• Hybrid Layer Design: Convolution + GQA: The 40-layer network consists of 30 “dual-gated short convolution blocks” and 10 layers of GQA. Short convolutions efficiently extract local sequence features, while attention layers model long-range dependencies. Alternating these components enables the model to maintain linear-layer efficiency while retaining strong long-context capabilities—the structural foundation for high speed and low memory usage on edge devices.
• In-Place Vocabulary Expansion & Long Context Training: Building upon the core architecture of LFM2-24B-A2B, the vocabulary was expanded from 65K to 128K via in-place tokenizer extension, significantly improving token efficiency for non-Latin languages like Chinese and Japanese. The context window was extended from 32K to 131K, combined with 300B-token specialized mid-training focused on long-form text, enabling stable handling of long sequences.
• Three-Stage Training Pipeline: Pre-training consumed approximately 33T tokens; mid-training first enhanced reasoning, math, code, and instruction-following abilities using 2T high-quality tokens at 32K context, then focused on long-text processing using 300B tokens at 131K context, both stages incorporating tool calling and multi-turn interaction data. Post-training followed a five-stage pipeline: SFT → Teacher Specialization → Dual-domain distillation (non-Agent and Agent) → Reinforcement learning within real-world agent frameworks like OpenClaw and Hermes, enabling the model to genuinely learn the “think-call-validate” loop in practical tool environments.

Follow WeChat and reply “open source” to join the AI Open Source Project Community

How to Use LFM2.5-24B-A2B

• Model Acquisition: Download LiquidAI/LFM2.5-24B-A2B from Hugging Face, selecting the appropriate version (native BF16, FP8, GGUF, ONNX, or MLX-4bit) based on your hardware.
• Environment Setup: Ensure Python environment includes transformers>=5.0.0, or deploy inference frameworks such as vLLM, SGLang, llama.cpp, or LM Studio.
• Model Loading: Load the model using AutoModelForCausalLM.from_pretrained() with bfloat16 precision.
• Prompt Construction: Format conversations automatically using tokenizer.apply_chat_template() with ChatML template.
• Parameter Configuration: Use official recommended settings: temperature=0.1, top_k=50, repetition_penalty=1.05. For chain-of-thought tasks, set max_new_tokens above 4096.
• Tool Calling Configuration: Define the tool list in JSON format within the system prompt. The model will output function calls marked with <|tool_call_start|>, returning results under the tool role after execution.
• Integration with Agent Frameworks: Launch a local service via OpenAI-compatible API, then integrate into agent frameworks such as OpenClaw, Hermes Agent, Pi, or OpenCode.

Core Advantages of LFM2.5-24B-A2B

• Extreme Efficiency: Only 2.45B parameters activated out of 24B total; achieves 184 tok/s on Apple M5 Max—1.7× faster than the fastest comparable MoE model—with minimal memory footprint.
• Small but Mighty: Matches performance of Qwen3.6-35B-A3B (50% larger parameters) on Agent and tool benchmarks: LiveClaw 48.09 vs. 38.19, AutomationBench surpassing second place by 3×.
• End-to-End Inference: Pure chain-of-thought model that reasons before answering, with each inference token being highly cost-efficient due to low activation overhead.
• Native Agent Capabilities: Released with Agentic RL training completed natively in mainstream frameworks like OpenClaw, Hermes, Pi, and OpenCode—tool calling ready out of the box.
• Open and Unrestricted: Both base and post-training models have openly available weights, enabling free download, fine-tuning, and commercial use.
• Full Platform Coverage: Native support for llama.cpp, MLX, vLLM, SGLang, ONNX; compatible with Apple, AMD, Qualcomm, and NVIDIA hardware.

Project Repository for LFM2.5-24B-A2B

• Hugging Face Model Hub: https://huggingface.co/LiquidAI/LFM2.5-24B-A2B

Competitive Comparison with Similar Models

Application Scenarios for LFM2.5-24B-A2B

• Edge-Side Private Agent: Run local intelligent agents continuously on personal devices like MacBook, performing planning, tool calling, and task execution—all data remaining strictly within-device.
• Offline/Edge Assistant: Deployed via GGUF, ONNX, or MLX formats onto unstable or offline environments (e.g., automotive systems, remote field operations, embedded terminals), providing reliable offline inference.
• Smart Tool Orchestrator: Acts as the driving model for frameworks such as OpenClaw, Hermes Agent, and OpenCode, executing multi-step workflows like scheduled report generation, file organization, and code operations.
• Business Process Automation: Leverages strengths in ToolSandbox and τ³-Bench Banking to handle business interactions requiring system queries and multi-round confirmation—such as banking, recruitment, and order processing.

Source: AI Tools Collection