Ling 3.0 Tiny

Ling 3.0 Tiny is a compact Mixture-of-Experts (MoE) language model developed by InclusionAI. Designed for resource-constrained environments and high-throughput agentic workflows, the model contains approximately 7.9 billion total parameters, with only 1.3 billion parameters activated per token. This sparse activation allows the model to provide reasoning-heavy performance while maintaining the speed and low latency profile typical of smaller models.

Architecture and Design

The model utilizes a native hybrid-linear architecture that integrates Kimi Delta Attention (KDA) with Multi-Head Latent Attention (MLA). This hybrid configuration is engineered to balance high-precision transformer accuracy with linear-scaling efficiency, which is particularly effective for processing long-context inputs. By alternating gated KDA layers with sparse MoE blocks, Ling 3.0 Tiny achieves stable performance in long-horizon tasks while minimizing the overall computational footprint.

Key Capabilities

Ling 3.0 Tiny supports dual execution modes: Thinking mode for complex, multi-step logical reasoning and Instant mode for low-latency, immediate responses. It natively supports function calling, structured outputs, and prompt caching. These features make it suitable for production-grade AI agents that require reliable instruction following and precise tool usage during multi-turn interactions.

The model features a native context window of 262,144 tokens, enabling it to handle extensive documentation, large datasets, or deep code repositories without requiring frequent recomputation of context. This long-context capacity, paired with a 32,768-token maximum output, supports complex agentic workflows such as document intelligence and long-horizon research tasks.