Ling 3.0 Flash scores 38 points on the Artificial Analysis Intelligence Index, marking a significant rise from its predecessor and placing it on par with Qwen3.6 27B while using far fewer active parameters. This performance makes it the smartest open model under 124 billion total parameters, though it still trails the leader, DeepSeek V4 Flash, which sits at 52 points. The model also shows marked improvements in reliability, with hallucination rates dropping from 97 to 44 percent on the AA Omniscience test. It now refuses to answer questions without reliable data far more often and demonstrates strong gains in agentic tasks, including on the t3-Bench Banking benchmark.
Ant Group’s inclusionAI releases the model under an MIT license via the inclusionAI API and through DeepInfra, with weights available directly on Hugging Face. On per-token pricing, Ling 3.0 Flash beats every comparably capable model, even though it burns through more tokens on complex tasks than similarly strong alternatives. It remains cheaper than Qwen3.6 27B on a per-task basis.
- Available under MIT license
- Weights hosted on Hugging Face
- Lower hallucination rate than previous version
