On the day Jensen Huang, CEO of NVIDIA, supported open artificial intelligence models in his first post on X alongside 25 industry signatories, the Ant Group Bailing division unveiled the Ling-3.0-flash model. This model is an executing model with a total of 124 billion parameters but activates only 5.1 billion parameters, demonstrating higher intelligence density than its trillion-parameter flagship model.
Ling-3.0-flash showed high results across 34 evaluation areas, ranking first in 15 cases and second in 19. The average score was 67.6, equating it to DeepSeek V4 Flash. Furthermore, the model outperformed its own trillion-parameter Ring-2.6-1T model by almost 8 points. The model achieves an intelligence density of 0.5 and an efficiency coefficient of 13.2, which are the highest among comparable models, proving that targeted design of an executing model can surpass general scaling when handling specific tasks.
The model is designed as a strategic bet on the paradigm of executing models, distinct from reasoning models. Ling-3.0-flash focuses on speed and efficiency for real-time agent tasks, coding, and function calling. For instance, in the SWE-Bench Pro test for fixing errors in code repositories, it scored 56.6% in first place. In ArtifactsBench, designed for creating complete interactive components in a single pass, it achieved 77.0%, surpassing second place by more than 10 points. In MiniAppBench, which requires full application generation from a single instruction, it reached 25.3% compared to an average of 17% across 16 models.
BFCL-v4 function calling accuracy reached 73.0%, matching Claude-Sonnet-4.6 for first place. In the end-to-end agent evaluation GDPVal v2-AA, the model scored 1107 points, the highest among tested models. The model natively supports context windows of 256K, where MRCR_256K achieved 81.1%. Multi-step instruction following Multi-IF reached 87.7% in second place. In the LIFEBench long-term memory instruction test, it ranked first with a result of 77.3%. The WideSearch agent search capability reached 73.6% in third place, and in multi-agent collaboration mode, BrowseComp scored 82.0%, nearly matching Claude-Sonnet-4.6 (82.1%).
The full suite of agent-oriented evaluations positions Ling-3.0-flash as optimized for the autonomous task execution paradigm, where models use tools, navigate interfaces, and accomplish multi-stage goals, rather than generating long chains of reasoning. Ant Bailing has made the model open, and weights will be available after the free API trial period ends in August. The release, simultaneous with a letter supported by NVIDIA, signals that Chinese AI companies continue to embrace open source despite Silicon Valley discussions about its risks. Ling-3.0-flash emphasizes efficiency as a key differentiator: instead of competing on total parameter count, the model demonstrates that for agent execution workloads, a well-designed compact executing model with targeted training can outperform models eight times larger. This aligns with industry trends toward task-specific models, as inference deployment economics become the primary constraint. The weekly ranking on OpenRouter rose from 9th to 6th place within five days, while July traffic volume was only 0.5% lower than DeepSeek V4 Pro, confirming market demand for efficient executing models.
Artificial intelligence has become one of China's new export directions, and the country is leading this race by developing more efficient, low-cost, and open-source models.
There is a trend where American companies are beginning to switch to Chinese AI solutions. Furthermore, the rest of the world is moving in the same direction.
