Ling 3.0 Tiny

Ling 3.0 Tiny is a compact MIT-licensed reasoning checkpoint with 7.9B total parameters and 1.3B active per token.

Local reasoning and agents

Tiny uses a hybrid KDA and MLA attention architecture with sparse experts. It supports ordinary responses and configurable thinking through enable_thinking. The publisher documents local deployment on systems including Apple Silicon and DGX Spark, making it a candidate for a coding assistant or small tool agent whose inference stays on your hardware.

Context and memory

The checkpoint configuration sets 131,072 positions. A separate serving recipe extends to 256K through YaRN; enabling that recipe is a deployment choice. The active parameter count describes token computation, while memory must still hold the complete model, caches and runtime state. The publisher’s small-context memory measurement should not be treated as a guarantee at maximum context.

A useful local trial

Start with a short repository task, a tool call with a known result and an instruction that requires looking back at earlier history. Compare thinking enabled and disabled using accepted results and total elapsed time. A concise answer is useful only if it preserves the task’s constraints.

Deployment and licence

Use a runtime with support for the model’s hybrid architecture and chat template. Confirm parser behaviour before attaching tools, and validate returned arguments in the application. The weights are MIT-licensed; the checkpoint does not establish a universal hosted token price.

Sources

Model documentation