Liquid AI Releases LFM2.5-24B-A2B: Fast On-Device Reasoning Agent

TL;DR: Liquid AI has released LFM2.5-24B-A2B, a 24B-parameter reasoning model with only 2B active parameters designed for running fast, private agents on local devices. The model achieves 184 tokens/second on Apple M5 Max and over 21K tokens/second on NVIDIA B300 GPU, while remaining competitive with larger models on agentic and tool-use benchmarks. Both the base and fine-tuned versions are available as open-weight models on Hugging Face.

What's New

Liquid AI has launched LFM2.5-24B-A2B, an open-weight reasoning model built specifically for on-device agent workflows. Published on October 13, 2026, the model features 24 billion total parameters but uses only 2 billion active parameters per token through its Mixture-of-Experts (MoE) architecture.

The model builds on its predecessor LFM2-24B-A2B with significant upgrades: an expanded context window from 32K to 131K tokens, doubled vocabulary from 65K to 128K tokens for improved non-Latin language support, and dramatically increased pre-training from approximately 17 trillion to 33 trillion tokens. Unlike earlier versions, LFM2.5-24B-A2B is a reasoning-only model that generates explicit chains of thought before producing final answers.

The architecture features 40 layers with 64 experts per MoE block using top-4 routing, combined with gated short convolutions and grouped-query attention (GQA). Training involved five post-training stages including supervised fine-tuning, teacher specialization, multi-domain on-policy distillation for both non-agentic and agentic tasks, and agentic reinforcement learning within popular agent harnesses like OpenClaw, Hermes Agent, Pi, and OpenCode.

On benchmarks, LFM2.5-24B-A2B scored 48.09 on LiveClaw compared to Qwen3.6-35B-A3B's 38.19, despite having approximately 50% fewer total and active parameters. The model leads on five of eight tool-use and agentic benchmarks including AutomationBench (27.69 versus next-best 8.72), BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking. It also tops all three instruction-following benchmarks tested.

Pricing and Availability

Both LFM2.5-24B-A2B and LFM2.5-24B-A2B-Base are available as open-weight models on Hugging Face. No pricing information was provided, as these are downloadable models rather than API services. Users can run the models on their own hardware without ongoing costs.

The models ship with day-one support across major inference frameworks including llama.cpp, MLX, vLLM, SGLang, and ONNX, enabling deployment across various hardware configurations.

How to Try It

Developers can start using LFM2.5-24B-A2B by installing transformers version 5.0.0 or later with pip install -U transformers. The model loads through standard Hugging Face interfaces using the model ID "LiquidAI/LFM2.5-24B-A2B".

For agent workflows, users can serve the model locally behind any OpenAI-compatible backend and configure their agent harness to connect to it. For example, with Hermes Agent, users set the provider to custom, point the base URL to their local server (port varies by backend: llama.cpp and MLX use 8080, vLLM uses 8000, SGLang uses 30000), and specify the model name and 131,072-token context length.

Liquid AI provides configuration examples for OpenClaw and Pi in the model card, along with a complete setup guide. Fine-tuning tutorials are also available for adapting the model to specific tasks.

Who It's For

LFM2.5-24B-A2B targets developers building agents that need to reason through multi-step tasks, call tools, and run reliably on local hardware for privacy or cost reasons. The model's efficient 2B active parameters make continuous agent operation feasible on consumer hardware.

The model particularly suits applications requiring strong tool-use and instruction-following capabilities. However, Liquid AI notes that larger models maintain advantages on knowledge-heavy benchmarks like GPQA Diamond, HLE, and AA-Omniscience, as well as coding benchmarks like SWE-Bench Verified and SciCode. For knowledge-intensive workloads, the company recommends pairing LFM2.5-24B-A2B with retrieval systems or web search.

Alternatives

Competing models evaluated against LFM2.5-24B-A2B include Qwen3.6-35B-A3B (35B total, 3B active parameters), Gemma-4-26B-A4B-it (26B total, 4B active parameters), and Nemotron-3.5-Lightning (30B total, 3B active parameters). All these alternatives feature more total and active parameters than LFM2.5-24B-A2B, yet the Liquid AI model remains competitive or superior on many agentic benchmarks while offering faster inference speeds on both CPU and GPU hardware.