AntLing Releases Ling-3.0-Based Models

AntLing has open-sourced a total of six base model checkpoints for Ling-3.0-tiny and Ling-3.0-flash, covering pre-training, mid-training, and WSM merging stages. These models have not undergone post-training, providing researchers with a flexible starting point for additional pre-training or fine-tuning.

A key feature is the adoption of WSM (Weighted Checkpoint Merging) to replace learning rate (LR) decay, enhancing the continuity of the training process. This allows for exploring various LR decay strategies offline, enabling validation of strategies on tiny-base with the same training recipe before scaling to flash-base.

• Ling-3.0-tiny-base: 7.9B total parameters (1.3B active). It demonstrates equal or superior performance on coding benchmarks compared to previous-generation models, with half the parameter count.
• Ling-3.0-flash-base: 124B total parameters (5.1B active). It delivers performance comparable to models 2–3 times larger on coding, reasoning, and long-context tasks, making it suitable for adaptation to specialized domains such as finance and healthcare.