Ling 3.0 Tiny (huggingface.co)

Ling-3.0-tiny has been launched as a lightweight hybrid reasoning model, comprising 7.9 billion total parameters, but activating only 1.3 billion parameters per token. This innovative model is optimized for low inference costs, making it ideal for deployment on local and resource-constrained hardware while retaining advanced reasoning and agentic capabilities. The use of BF16, FP8, and INT4 weight formats broadens its application across diverse hardware settings. Significantly, Ling-3.0-tiny integrates a unique hybrid-linear architecture that employs a 3:1 stacking of Kimi Delta Attention (KDA) and Multi-Head Latent Attention (MLA) layers, ensuring efficient processing of long contexts. The model’s native hybrid reasoning allows it to handle both speedy responses and complex multi-step reasoning tasks. Performance benchmarks show it achieves speeds over 160 tokens/second, with substantial efficiency relative to its parameter footprint. Such advancements enhance the feasibility of deploying sophisticated AI models in everyday applications, particularly in environments without access to high-end computational resources.