Liquid AI Unveils 8B-A1B MoE Model Trained on 38T Tokens

Liquid AI has launched LFM2.5-8B-A1B, an edge model capable of fast and reliable tool calling on consumer hardware.

This model features an expanded context window of 128K compared to the previous version, and dramatically increased the pretraining data scale from 12T to 38T tokens. It also doubled the vocabulary size to 128K to improve tokenization efficiency for non-Latin languages.

In particular, this version is designed as a reasoning-only model. By having the model generate an explicit Chain of Thought before producing a final answer, it significantly improved quality while maintaining the computational efficiency of the MoE model.

Key features are as follows:

• On-device personal assistant: Optimized for complex instruction following and tool calling chaining
• Compressed performance: Competitive performance against larger dense and MoE models
• Top-tier throughput: Best-in-class speed for both CPU and GPU inference (supports llama.cpp, MLX, vLLM, SGLang)

Benchmark results show dramatic performance improvements over the previous model on key metrics including AA-Omniscience Index, IFEval, and MATH500.