AntLing drops Ling-3.0's full six-checkpoint matrix
Six public MIT-licensed base checkpoints for Ling-3.0 — pretrain, mid-train, WSM-merged in tiny and flash flavors
AntLing has open-sourced the complete six-checkpoint matrix for its new Ling-3.0 base model, giving researchers and engineers direct access to two model sizes (tiny and flash) at three distinct training stages: pretrained, mid-trained, and WSM-merged. Each checkpoint is published in a separate official GitHub repository with an MIT license, and all were ungated at launch. Critically, these are base checkpoints only—they have not been post-trained, instruction-tuned, or wrapped into end-user chat interfaces.
The release is designed for builders who need to continue pretraining, fine-tune for downstream tasks, or experiment with stage-specific optimization. AntLing has provided a first-party stage map to help users decide where to enter the training pipeline, but the company is not endorsing a single “best” stage across all tasks. The open artifacts also leave room for community-driven validation on quantization behavior and downstream performance once the checkpoints are adapted.
- Six public MIT-licensed Ling-3.0 base checkpoints: tiny and flash in pretrained, mid-trained, and WSM-merged variants
- All repositories are ungated and intended for continued pretraining, fine-tuning, and research—not ready-to-use chat models
- Builders can select their entry point on the training trail using AntLing’s first-party stage map
Why It Matters
Open-sourcing the full checkpoint family accelerates base-model research and allows teams to tailor Ling-3.0 to specialized domains