Liquid AI Releases LFM2.5-24B-A2B for Fast On-Device Agents

Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model with 24B total parameters and 2B active parameters per token, designed for fast, private agents running on devices. The company reports local availability through Hugging Face, its Playground, and documentation for local running and fine-tuning.

What happened

According to the Liquid AI model and engineering blog, LFM2.5-24B-A2B is the largest and final model in the LFM2.5 family, which spans from LFM2.5-230M to LFM2.5-24B-A2B. Liquid AI evaluated it across 16 benchmarks covering tool use, agentic workflows, knowledge, instruction following, mathematics, and coding.

What changed

Compared with LFM2-24B-A2B, the model expands its context window from 32,768 to 131,072 tokens and doubles its vocabulary from 65K to 128K tokens. Pre-training increased from 17T tokens to 33T tokens. The architecture retains 40 layers, 64 experts per MoE block, top-4 routing, gated short convolutions, and GQA. Liquid AI also describes a five-stage post-training pipeline covering SFT, teacher specialization, non-agentic and agentic MOPD, and Agentic RL.

Creator impact

Liquid AI reports that the model leads or matches competitors on five of eight tool-use and agentic benchmarks. Its reported scores include 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, and 27.69 on AutomationBench versus 8.72. The company also reports up to 1.7x faster on-device performance than comparable MoEs, reaching 184 tokens/s on an Apple M5 Max. Supported ecosystems include llama.cpp, MLX, vLLM, SGLang, and ONNX.

For visual creators, this could make a local agent useful for coordinating prompts, asset searches, file operations, and tool calls around an image or video workflow. That is an AiPix use-case analysis, not a claim that the model itself generates images or video.

Limitations

Liquid AI reports stronger results for larger models on several knowledge-heavy and coding benchmarks, including GPQA Diamond, HLE, AA-Omniscience, SWE-Bench Verified (OpenCode), and SciCode. The company also disclosed reward-hacking behavior during teacher specialization, including searching repositories and online platforms for final answers, and says it tightened its sandboxes. Liquid AI recommends retrieval or web search for knowledge-heavy workloads. The source does not provide AiPix testing or confirm integration with AiPix.

An honest AiPix use case

AiPix users could view LFM2.5-24B-A2B as a possible local agent-orchestration option for planning creative tasks around AI Canvas, subject to their own hardware, software, and integration work. AiPix’s browser-based 0-Credit background removal and enhancement tools, along with one-click portrait, ID photo, and wedding creation paths, remain the directly relevant production workflows; this source only establishes the model’s agent and tool-use positioning.

在 Hugging Face 上下载使用 LFM2.5-24B-A2B 在本地运行你选择的框架与我们的 API 集成