Ling-3.0-flash-VL, Ant Group’s new flash tier open weights model, scores 25 on the Artificial Analysis Intelligence Index. At 124B total parameters 5.5B active parameters, it sits on the Intelligence vs. Active Parameter Pareto Frontier

@AntLingAGI has released Ling-3.0-flash-VL, an open weights reasoning model that adds image and video understanding to Ling-3.0-flash. Its mixture-of-experts architecture activates 5.5B of its 124B parameters per token, with support for a 256K token context window.

Key results:

➤ Ling-3.0-flash-VL sits on the Pareto Frontier for Intelligence vs. Active Parameters, scoring 25 with 5.5B active parameters. Among models with a similar total size, Qwen3.5 122B A10B (Reasoning) scores 16 and Mistral Medium 3.5 (high) scores 15.

➤ Ling-3.0-flash-VL features lower hallucination rate than comparable models, but with limited factual recall. Ling-3.0-flash-VL scores 14% on AA-Omniscience Accuracy and 22% on Hallucination Rate. Inkling Small answers more questions correctly at 33% Accuracy, but has a much higher Hallucination Rate at 63%.

➤ There remains room for improvement for difficult agentic tasks for Ling-3.0-flash-VL. The model scores 16% on AutomationBench-AA, which tests completing workflows across business apps while respecting guardrails, and 0% on Terminal-Bench v4.0, a harder terminal-use benchmark.

➤ Ling-3.0-flash-VL is decently verbose with its output. The model averages ~50k output tokens per Intelligence Index task vs. ~30k for Inkling Small (Reasoning), despite the similar overall scores. This will have cost implications for workloads with a high level of reasoning.

Additional model details:

➤ Type: Open weights reasoning model.

➤ Size: 124B total parameters, 5.5B active per token (MoE).

➤ Context window: 256K tokens.

➤ Modalities: Text, image, and video input; text output.

➤ API availability: First- and third-party APIs.

➤ License: MIT.