From the source
Ant Ling has open-sourced Ling-3.0-tiny, a sparse model with 7.9 billion total parameters that activates 1.3 billion per token, to make tool use and agent execution practical on local hardware.
- The company says the models that handle function calling well are often expensive to run, which keeps always-on assistants off personal devices and pushes sensitive data toward remote APIs.
- Ling-3.0-tiny was released as weights in BF16, FP8 and INT4, and Ant Ling says it reaches 62.72 on BFCL v4 in its reported evaluations, alongside balanced results in general agent tasks, coding agents, reasoning, instruction following and long-context understanding.
- On the Artificial Analysis Intelligence Index v4.1.1 it scores 25, and 16 on the Artificial Analysis Agentic Index.
- The architecture pairs a 3:1 KDA-MLA attention stack with a sparse mixture-of-experts network of 128 routed experts, activating eight routed experts and one shared expert per token.
- Ant Ling says the model supports context lengths of up to 1 million tokens, though practical length depends on runtime, memory and deployment configuration; its SGLang recipe demonstrates a 256K configuration and the reported Mac memory measurement uses an 8K context.
