Liquid AI dropped LFM2.5-24B-A2B on Hugging Face this morning, October 13, 2026. This open-weight reasoning model features 24 billion total parameters but only 2 billion active per token. It is designed for local agentic workflows, automatically generating chain-of-thought reasoning before executing tool calls. Both the instruct and base versions are available for download immediately.

Architecture and Training Data

The model retains the 40-layer mixture-of-experts structure of its predecessor, LFM2-24B-A2B, utilizing 64 experts per block with top-4 routing. Liquid AI expanded the context window from 32K to 131K tokens and doubled the vocabulary to 128K tokens to improve handling of non-Latin text. Pre-training data grew significantly from approximately 17T to 33T tokens. Post-training involved agentic reinforcement learning within harnesses like OpenClaw, Hermes Agent, Pi, and OpenCode.

Benchmark Performance and Hardware Speed

Vendor-supplied benchmarks show strong results on automation tasks. LFM2.5-24B-A2B scored 27.69 on AutomationBench, significantly outpacing Qwen3.6-35B-A3B's 8.72. It also achieved 48.09 on LiveClaw and 96.75 on IFEval. However, larger models retain the lead on complex reasoning tasks; Qwen3.6-35B-A3B scored 83.69 on GPQA Diamond compared to LFM2.5's 74.42. On hardware, the model hits 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395. A single NVIDIA B300 can push over 21K output tokens per second at high concurrency.

Integration and Missing Details

Developers can serve the model locally behind any OpenAI-compatible backend. Liquid AI provided examples for Hermes Agent, emphasizing a 131,072-token context setting with tool-use enforcement enabled. The model supports llama.cpp, MLX, vLLM, SGLang, and ONNX out of the box. Fine-tuning tutorials are linked for task-specific adaptation. Notably, the release lacks explicit license terms, pricing information, and real-world production deployment metrics. The benchmarks reflect lab conditions, not enterprise CRM or helpdesk environments.

Key Takeaways

  • LFM2.5-24B-A2B offers 131K context and 2B active parameters for efficient local inference.
  • It outperforms Qwen3.6-35B-A3B on automation benchmarks but trails on complex reasoning.
  • Speed reaches 184 tokens per second on Apple M5 Max hardware.
  • License terms and production metrics are currently absent from the release notes.

The Bottom Line

This is a solid local agent for automation-heavy tasks, but don't sleep on the missing license details. Always verify commercial usage rights before integrating vendor-only benchmarks into your stack.