Ling 3.0 Tiny Leads Small Model Rankings on Artificial Analysis Benchmark
A new Artificial Analysis benchmark update shows InclusionAI's Ling 3.0 Tiny model leading performance among small language models. Despite activating only 1.3 billion parameters per token, it outperforms competing small-scale LLMs. This demonstrates the growing capability of sparse Mixture-of-Experts architectures to deliver high intelligence while minimizing active computational costs. Such efficient small models allow developers to run highly capable AI locally on resource-constrained devices at lower costs. Ling 3.0 Tiny is a 7.9-billion total parameter Mixture-of-Experts (MoE) model developed by InclusionAI, Ant Group's AI division. It features a native 256K-token context window, switchable reasoning modes, and prompt caching alongside native function calling.
## BACKGROUND
Artificial Analysis is an independent benchmarking platform that aggregates intelligence scores, latency, pricing, and speed metrics for large language models. Mixture-of-Experts (MoE) is an architecture where only a subset of total parameters (experts) is routed and activated per input token, allowing models to maintain high accuracy while drastically reducing compute requirements during inference.