LFM2.5-24B-A2B: High-Performance, Private AI Agent for On-Device Use

LFM2.5-24B-A2B is introduced as a state-of-the-art, highly efficient, and private AI agent designed specifically for reliable on-device operation. The model is engineered to excel in complex agentic workflows, tool utilization, and logical reasoning, positioning itself as a formidable competitor even against much larger and more complex models, such as Qwen3.6-35B-A3B, which boasts approximately 50% more total and active parameters. In head-to-head comparisons, LFM2.5-24B-A2B achieved a score of 48.09 on the LiveClaw benchmark, significantly outperforming Qwen3.6-35B-A3B's score of 38.19, demonstrating superior performance in practical, real-world agentic tasks.

One of the most notable architectural advancements is its approach to reasoning. Unlike traditional models that might integrate reasoning into their final output, LFM2.5-24B-A2B is designed as a dedicated reasoning-only model. This means that before generating its final answer, the model is forced to produce an explicit, step-by-step chain of thought. This structured reasoning process is crucial for reliability and transparency, allowing developers and users to trace the model's logic and understand how it arrived at a conclusion. Furthermore, the model maintains a highly efficient reasoning mechanism, utilizing only 2 billion active parameters per token, which keeps the computational cost of each reasoning step remarkably low.

In terms of technical efficiency, LFM2.5-24B-A2B showcases exceptional inference speeds across multiple hardware platforms. On Apple hardware, specifically the M5 Max, it achieves an impressive decoding rate of 184 tokens per second, making it the fastest Mixture-of-Experts (MoE) model among its peers. For GPU deployments, the model reaches over 21,000 output tokens per second on a single NVIDIA B300 GPU, solidifying its status as the fastest model in its size class. These high speeds, combined with its low memory footprint, make it ideal for deployment in resource-constrained, private, or edge computing environments, ensuring that complex AI capabilities do not require massive cloud infrastructure.

Architecturally, LFM2.5-24B-A2B builds upon its predecessor, LFM2-24B-A2B, retaining a robust structure of 40 layers and 64 experts per MoE block, utilizing top-4 routing. Significant upgrades include a massive expansion of the context window, increasing it from 32,000 tokens to 131,000 tokens. This extended context capacity allows the model to process and maintain coherence over extremely long documents, conversations, or codebases. Additionally, an in-place tokenizer expansion doubles the vocabulary size from 65,000 to 128,000 tokens, greatly enhancing its ability to handle non-Latin scripts and diverse global languages with improved tokenization accuracy.

The training regimen for LFM2.5-24B-A2B is exceptionally comprehensive, involving a massive scaling up of pre-training data from approximately 17 trillion tokens to 33 trillion tokens. This foundational training is supplemented by two critical mid-training phases. The first adds 2 trillion tokens within a 32K context window, specifically weighting the data toward reasoning, advanced mathematics, coding tasks, precise instruction following, and multilingual text. The second phase adds a substantial 300 billion tokens within the expanded 131K context, focusing on deep long-context data understanding. Crucially, throughout both mid-training phases, the model is exposed to tool calls and multi-turn conversational trajectories, preparing it for complex agentic behavior.

The transformation into a full-fledged agent is completed through a rigorous five-stage post-training process. This process begins with Multi-domain on-policy distillation (MOPD) for non-agentic tasks, distilling knowledge from various specialist teachers. This is followed by a second MOPD stage, which is specifically focused on refining agentic capabilities. The core of the agent training is the Agentic Reinforcement Learning (Agentic RL) phase. During this phase, the model is trained within specialized agent harnesses, including industry-leading frameworks such as OpenClaw, Hermes Agent, Pi, and OpenCode. This multi-faceted training ensures that the model is not only knowledgeable but also proficient in planning, executing, and correcting multi-step tasks using external tools.

When evaluated across 16 diverse benchmarks covering tool use, agentic workflows, general knowledge, instruction following, mathematics, and coding, LFM2.5-24B-A2B demonstrated remarkable competitive parity. The model leads or matches the competition on five out of eight tool-use and agentic benchmarks, including AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking. Its performance on AutomationBench is particularly outstanding, scoring more than three times higher than the next best model, confirming its robust ability to execute automated tasks. Furthermore, it leads on all three instruction-following benchmarks, solidifying its reliability for developers building complex applications. While acknowledging that larger models might retain an edge on highly knowledge-intensive benchmarks (like GPQA Diamond or AA-Omniscience) or specialized coding benchmarks (like SWE-Bench Verified), the developers strongly recommend pairing LFM2.5-24B-A2B with Retrieval-Augmented Generation (RAG) or web search capabilities for optimal performance in knowledge-heavy workloads. Overall, LFM2.5-24B-A2B is presented as the definitive foundation for building reliable, reasoning-capable, and highly efficient agents that can run autonomously on local devices.

For maximum developer accessibility, the model is open-weight and available on Hugging Face, with day-one support across the entire inference ecosystem. This includes major frameworks such as llama.cpp, MLX, vLLM, SGLang, and ONNX. This broad compatibility means that developers can integrate LFM2.5-24B-A2B into their existing toolchains with minimal friction. The model can be easily served locally behind any OpenAI-compatible backend, allowing agent harnesses like Hermes Agent or OpenClaw to point directly to it for private, secure operation. The availability of both the base model and the fine-tuned agent version (LFM2.5-24B-A2B-Base) ensures flexibility for both foundational research and immediate commercial application. The model also comes with comprehensive fine-tuning tutorials, enabling users to adapt its powerful capabilities to highly specific, niche tasks required by their unique enterprise applications. The combination of superior agentic performance, industry-leading inference speed on edge devices, and open-weight accessibility makes LFM2.5-24B-A2B a disruptive force in the local AI agent market.