Liquid AI Launches LFM2.5-24B-A2B: On-Device Reasoning Agent at 184 tok/s with 131K Context
By Ali Al-Montaser — Senior AI Technical Editor
Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model with 24B total and 2B active parameters, designed for on-device AI agents. It outperforms Qwen3.6-35B-A3B on agentic tasks by 50% in parameters, scoring 48.09 vs 38.19 on LiveClaw and hitting 184 tok/s on Apple M5 Max. Available immediately on Hugging Face with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX.
Executive Overview
Liquid AI has officially announced the launch of LFM2.5-24B-A2B, an open-weight reasoning model with 24 billion total parameters and only 2 billion active parameters, purpose-built for running AI agents locally on devices. The model is available immediately on Hugging Face in two versions: LFM2.5-24B-A2B and LFM2.5-24B-A2B-Base, with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. It delivers 184 tokens/second on Apple M5 Max and outperforms larger models like Qwen3.6-35B-A3B on agentic tasks, achieving 48.09 vs 38.19 on LiveClaw. This release marks a significant step forward for on-device AI, enabling private, low-latency agentic workflows without cloud dependency.
📊 Official Technical Specifications & Data Sheet
| Technical Aspect | Confirmed Official Data |
|---|---|
| 💰 Pricing & Usage Cost | Open-weight and free to download on Hugging Face. No official API fees; cost = local hardware or self-hosting. Low inference token cost due to only 2B active parameters. |
| 🌐 Platforms & Immediate Availability | Hugging Face (LFM2.5-24B-A2B and LFM2.5-24B-A2B-Base). Day-one support: llama.cpp, MLX, vLLM, SGLang, ONNX. Local ports: llama.cpp/MLX 8080, vLLM 8000, SGLang 30000, LM Studio 1234. Compatible with transformers>=5.0.0. |
| ⚡ Performance & Speed Benchmarks | 184 tokens/s on Apple M5 Max, 114 tokens/s on Ryzen AI Max+ 395 (1.7x faster than the next with lowest memory). Over 21,000 output tokens/s on a single NVIDIA B300 at high concurrency. LiveClaw: 48.09 vs 38.19 for Qwen3.6-35B-A3B. AutomationBench: 27.69 (more than 3x the next). |
| 🛡️ Security & Breach Resistance | Full on-device operation ensures data privacy and no external server transmission. No explicit security ratings in the official statement; the core security advantage is local isolation of agents. |
| 🧠 Context Window | 131,000 tokens (upgraded from 32K in the previous version). Mid-training: 2T tokens at 32K + 300B tokens at 131K. Vocabulary expanded from 65K to 128K tokens. |
| 🌍 Arabic & Regional Support | Enhanced support for non-Latin languages via vocabulary expansion (128K tokens) and multilingual training (2T tokens). Available globally for download and local use without geographic restrictions. |
Deep-Dive Features & Architecture
LFM2.5-24B-A2B is built on a Mixture of Experts (MoE) architecture with 40 layers and 64 experts per MoE block with top-4 routing, alongside a hybrid mix of gated short convolutions and GQA. The most notable upgrade is the expansion of the context window from 32K to 131K tokens, with in-place tokenizer expansion from 65K to 128K tokens, improving token efficiency for non-Latin scripts including Arabic. The model is a reasoning-only model that produces an explicit chain of thought before the final answer, unlike its predecessor LFM2-24B-A2B.
Pre-training was expanded from ~17T to ~33T tokens, followed by an additional 2T tokens at 32K context (weighted toward reasoning, math, code, instruction following, and multilingual text) and 300B tokens at 131K context (weighted toward long-context data). Post-training includes five stages: supervised fine-tuning (SFT), then non-agentic multi-domain on-policy distillation (MOPD), then agentic MOPD, then agentic reinforcement learning (Agentic RL) within agent environments such as OpenClaw, Hermes Agent, Pi, and OpenCode.
Benchmark & Competitive Performance
The model was evaluated against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning on 16 benchmarks. Despite being the smallest in the group, it outperforms or matches in 5 of 8 tool and agent benchmarks: AutomationBench (27.69 vs 8.72), BFCLv4 (65.01 vs 65.00), ToolSandbox (80.71 vs 75.02), LiveClaw (48.09 vs 38.19), and τ³-Bench Banking (8.04 vs 7.63). It also leads in all three instruction-following benchmarks: IFEval (96.75), IFBench (74.69), Multi-IF (84.38). In contrast, larger models retain superiority in knowledge benchmarks such as GPQA Diamond (74.42 vs 83.69) and HLE (7.40 vs 7.69) and in code such as SWE-Bench Verified (51.00 vs 66.20), so it is recommended to pair it with retrieval or web search for heavy knowledge workloads.
Industry Impact & Enterprise Adoption
The launch of LFM2.5-24B-A2B signals a growing trend toward efficient, on-device AI agents that prioritize privacy, latency, and cost. With only 2B active parameters, it enables persistent agentic workflows on consumer hardware like Apple M5 Max and Ryzen AI Max+ 395, reducing reliance on cloud APIs. Enterprises in finance, healthcare, and legal sectors can deploy local agents for sensitive data processing without external transmission. The model's strong performance on tool-use benchmarks makes it suitable for automation, customer support, and coding assistants. Its open-weight nature and day-one support across major inference frameworks lower the barrier for developers to integrate and customize. As edge AI continues to evolve, Liquid AI's focus on reasoning-only models with expanded context and multilingual support positions it as a key player in the local agent ecosystem.
Conclusion
LFM2.5-24B-A2B represents a significant advancement in on-device AI, combining a compact MoE architecture with a 131K context window and strong agentic capabilities. Its availability on Hugging Face with broad framework support empowers developers to build private, fast, and cost-effective AI agents. While it may not match larger models on knowledge-intensive tasks, its efficiency and tool-use prowess make it a compelling choice for edge deployment. As the industry shifts toward decentralized AI, Liquid AI's latest release sets a new benchmark for local reasoning agents.
Media Source: Hugging Face | Official Company Statement: Original Source | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
The model is open-weight and free to download and use locally on Hugging Face. There are no official API fees from Liquid AI; the only cost is local hardware (GPU/CPU) or self-hosting. This makes it a cost-effective option for persistent agents compared to paid API models.
Yes, the model supports non-Latin languages with improved efficiency thanks to in-place tokenizer expansion from 65K to 128K tokens. Mid-training included 2T multilingual tokens, enhancing token efficiency for Arabic text compared to the previous LFM2-24B-A2B version.
It is available immediately on Hugging Face in two versions: LFM2.5-24B-A2B and LFM2.5-24B-A2B-Base. It supports inference via llama.cpp, MLX, vLLM, SGLang, and ONNX with day-one support. It can run locally behind any OpenAI-compatible interface (ports: llama.cpp/MLX 8080, vLLM 8000, SGLang 30000, LM Studio 1234).
The model achieves 184 tokens/second on Apple M5 Max and 114 tokens/second on Ryzen AI Max+ 395, up to 1.7x faster than the next model in comparison with lower memory usage. On a single NVIDIA B300 GPU, it reaches over 21,000 output tokens/second at high concurrency.
It outperforms Qwen3.6-35B-A3B on LiveClaw with a score of 48.09 vs 38.19, and on AutomationBench with 27.69 vs 8.72 (more than 3x). It matches or exceeds in 5 out of 8 tool and agent benchmarks despite being about 50% smaller in total and active parameters.
AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
