LFM2.5-24B-A2B

LFM2.5-24B-A2B is Liquid AI's open-weight, text-only reasoning model for agents that run on local hardware. It has 24 billion total parameters but activates 2.45 billion per token through a 64-expert, top-4 MoE design. The official card positions it as the largest and final LFM2.5 model, with a 131,072-token context window and native tool calling.

What it is for

This is a practical choice when an agent needs to reason, call tools, and keep data on a device or self-managed server. Liquid publishes native, FP8, GGUF, ONNX, and MLX variants, so the same model can be evaluated with Transformers, vLLM, SGLang, llama.cpp, LM Studio, or Apple Silicon MLX. The weights use Liquid AI's lfm1.0 license, not an OSI open-source license, so teams should read its terms before distribution or commercial deployment.

Key capabilities

• Sparse reasoning model: 24B total parameters, 2.45B active parameters, 40 layers, and 64 experts per MoE block.
• Long-context agents: 131,072-token context and a 128,000-token vocabulary for multi-step tool workflows.
• Function calling: The model card documents Python-style tool calls by default and JSON calls when instructed.
• Local deployment formats: GGUF targets llama.cpp-compatible runtimes; ONNX targets cross-platform acceleration; MLX targets Apple Silicon.
• Multilingual input: The published language list includes English, Chinese, Japanese, Arabic, French, German, Italian, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, and Polish.

A sensible evaluation workflow

Start with the official chat template and an OpenAI-compatible local server. Give the model a small, typed tool set, log every call and result, and replay a held-out task suite. Use retrieval or web search for knowledge-heavy questions: Liquid explicitly recommends that pairing instead of treating the model as a closed knowledge base. Compare its latency, tool-call validity, context use, and recovery from failed tools against LFM 2.5-2.6B or another local model on your own hardware.

Limits and alternatives

Vendor benchmark results and throughput figures are useful leads, not deployment guarantees. Actual speed depends on quantization, runtime, memory bandwidth, prompt length, and tool-loop design. The model is text-only, and its license requires separate review. If a smaller always-on agent is enough, consider LFM 2.5-2.6B; if you need a hosted coding workflow, compare Codex CLI.

Getting started

Read the official model card, choose a supported format, and run a local tool-use evaluation before connecting it to production actions.