LFM2.5-8B-A1B logo

MODELS

LFM2.5-8B-A1B

by Liquid AI

modelLiquid AI
Visit WebsiteDocumentationGitHub

Overview

Liquid AI compact MoE reasoning model for on-device tool calling, function calling, instruction following, and agentic tasks.

Details

LFM2.5-8B-A1B is a Liquid AI text-only reasoning-tuned Mixture-of-Experts model. Official Liquid docs identify it as an 8B model with 1.5B active parameters and a 128K context for on-device tool calling and agentic tasks. The Hugging Face model card lists 8.3B total parameters, tool use, multilingual support, and inference support through Transformers, vLLM, SGLang, llama.cpp, ONNX, MLX, and GGUF variants. Liquid’s launch blog describes it as an edge model for fast, reliable tool calling on consumer hardware, with 38T-token pretraining and reinforcement learning.

When to Use

Use when evaluating a compact Liquid AI MoE reasoning model for on-device or edge-oriented agentic tasks. Use when a workflow needs tool calling or function calling with JSON tool definitions and tool-role results. Use when a text-only model with 128K context and multilingual support is required. Use when deployment targets include runtimes such as Transformers vLLM SGLang llama.cpp ONNX MLX or GGUF variants.

Getting Started

  1. Read the Liquid AI launch blog for the model overview and positioning.
  2. Review the Liquid text-model documentation for the catalog entry and model specifications.
  3. Use the Liquid tool-use documentation to implement function calling with JSON tool definitions
  4. special tool-call tokens
  5. tool-role results
  6. and final response regeneration.
  7. Open the official Hugging Face model card for weights/model-card details and supported inference/runtime options.
  8. Check Liquid AI pricing terms before commercial or organization-wide deployment.

Key Features

  • •Mixture-of-Experts reasoning model with 8B class size and 1.5B active parameters in Liquid docs.
  • •128K context window listed in Liquid docs and Hugging Face metadata.
  • •Text-only reasoning-tuned model listed on the official Hugging Face model card.
  • •Tool use and function-calling support via Liquid’s documented tool-use format.
  • •Inference support listed for Transformers
  • •vLLM
  • •SGLang
  • •llama.cpp
  • •ONNX
  • •MLX
  • •and GGUF variants.
  • •Positioned by Liquid AI as an edge model for consumer hardware.

Capabilities

  • •reasoning
  • •tool calling
  • •function calling
  • •instruction following
  • •multilingual text
  • •long-context text processing
  • •on-device inference
  • •agentic workflows

Last updated Jul 1, 2026