Liquid AI reveals 8B-A1B MoE trained on 38T (www.liquid.ai)

Liquid AI has unveiled its latest model, LFM2.5-8B-A1B, purpose-built for efficient tool calling on everyday consumer hardware. This new edge model significantly enhances its predecessor by expanding the context window from 32K to an impressive 128K tokens and scaling pretraining from 12 trillion to 38 trillion tokens. Additionally, it doubles the vocabulary to 128K to improve tokenization efficiency for non-Latin languages, yielding notable performance gains, especially in languages like Hindi and Arabic. By focusing on reasoning capabilities, the model delivers explicit chains of thought, bolstering its accuracy in instruction following and reducing hallucination rates, positioning it competitively against larger models. For the AI/ML community, LFM2.5-8B-A1B represents a leap toward accessible on-device AI, enabling real-time, private interactions without relying on cloud infrastructure. Its architecture uses mixture of experts (MoE) and advanced training methods to ensure high throughput and performance on both CPU and GPU, making it the fastest model in its class for consumer hardware. With day-one support across various inference platforms and having proven its reliability through various benchmarks, this model sets the stage for a new era of personal assistants capable of sophisticated operation directly on user devices, thereby championing both speed and privacy in AI deployments.