Liquid AI Releases LFM2.5-8B-A1B: An On-Device MoE Model With 8.3B Total and 1.5B Active Parameters

2026-05-28 · Source: Machine Learning ML & Generative AI News · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, quick

Summary

Liquid AI released LFM2.5-8B-A1B, an on-device Mixture-of-Experts (MoE) model, which activates only 1.5B of its 8.3B total parameters per token. This version introduces reasoning-only capabilities, producing an explicit chain of thought, and demonstrates a significant reduction in hallucination rates, with the Non-Hallucination Rate improving from 7.46 to 63.47, IFEval from 79.44 to 91.84, MATH500 from 74.80 to 88.76, and Tau² Telecom from 13.60 to 88.07. The model runs efficiently on various hardware, achieving 253 tok/s on an M5 Max under 6 GB, ~30 tok/s on a phone, and 18.5K tok/s on a single H100. It emphasizes tool calling, exemplified by the LocalCowork demo running 67 tools across 13 MCP servers on one laptop without cloud dependencies. The model offers open weights and day-one support for llama.cpp, MLX, vLLM, and SGLang.

Key takeaway

For AI engineers building local agents or on-device AI applications, you should consider LFM2.5-8B-A1B. Its efficient MoE architecture, combined with significantly reduced hallucination and robust tool-calling capabilities, offers a compelling solution for privacy-preserving, high-performance local AI. Evaluate its integration to develop agents that operate entirely on-device, minimizing cloud dependencies and data egress.

Key insights

LFM2.5-8B-A1B is an on-device MoE model that significantly reduces hallucinations and enables complex tool-calling.

Principles

• MoE architecture enables cheap reasoning tokens.
• Targeted RL reward can train models to abstain on unknown questions.

Method

A targeted avg@k RL reward trains the model to abstain on questions beyond its knowledge, significantly reducing hallucination rates.

In practice

• Run 67 tools across 13 MCP servers on one laptop.
• Utilize llama.cpp, MLX, vLLM, SGLang for deployment.

Topics

• Mixture-of-Experts
• On-device AI
• Hallucination Reduction
• Tool Calling
• Local Agents
• Model Inference

Best for: AI Architect, NLP Engineer, Research Scientist, AI Engineer, Machine Learning Engineer, AI Scientist

Related on AIssential

• Latest open artifacts (#18): Arcee's 400B MoE, LiquidAI's underrated 1B model, new Kimi, and anticipation of a busy month — Smaller open models are achieving competitive performance against larger counterparts through extensive pretraining and architectural innovations.
• z.ai's open source GLM-5 achieves record low hallucination rate and leverages new RL 'slime' technique — GLM-5 sets new benchmarks for open-source LLMs in hallucination reduction and agentic document generation at a competitive price.
• Qwen3.5 9B MoQ: Inside a Strong 3.6-bit GGUF — Advanced quantization and MoE architectures significantly boost LLM efficiency and performance on diverse hardware.
• Creating highly efficient agents: 450M tool-calling tokens distilled for post-training from top open-source models — Distilling large model capabilities into smaller, efficient agents enables advanced tool-calling on lightweight compute.
• 15B Active MoE BEATS OPUS 4.6 in Reasoning — Myo v2's superior reasoning stems from hybrid attention, MOPD, and MoE stabilization techniques.
• Overcoming LLM hallucinations in regulated industries: Artificial Genius’s deterministic models on Amazon Nova — A hybrid AI architecture enables deterministic, non-hallucinating LLM outputs for regulated industries by using generative models non-generatively.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning ML & Generative AI News.