Ant Group's InclusionAI Open-Sources Ling-3.0 Base Model Checkpoints Across Training Stages

InclusionAI (Ant Group's Ling team) open-sourced the Ling-3.0 base model family on Hugging Face around August 19, releasing 6 MIT-licensed checkpoints across pre-training, mid-training, and WSM merging stages for both tiny and flash sizes.

Confirmed

• The release includes Ling-3.0-tiny-base (7.9B) and Ling-3.0-flash-base (124B), with 6 checkpoints covering pre-training, mid-training, and WSM merging stages.
• All checkpoints are MIT-licensed base models without post-training, not intended as ready-to-use chat assistants but suited for continued pre-training.
• The training pipeline introduces WSM (Weighted State Merging), replacing conventional LR decay with a merging approach applied across the training cycle.
• Per the release framing, developers and researchers can pick any stage to continue training or compare across stages to study the Ling family's full evolution.

Why it matters

• Releasing weights across all training stages is rare in open source, enabling research into training dynamics and intermediate representations.
• Open-sourcing WSM as an LR-decay alternative gives the community directly verifiable material for training-method research.
• The MIT license plus dual sizes (7.9B and 124B) serve both academic experimentation and large-scale continued training.

2026-08-19 ~ 2026-08-21 · 6 related posts

• Ling-3.0 Open Source: 6 Base Checkpoints Released with WSM Training Method — bclavie · 2026-08-19
• [source] InclusionAI Releases Ling-3.0-tiny-base Model — jacek2023 · 2026-08-19
• AntLing open-sources Ling-3.0 models using WSM to replace LR decay — AcanthisittaOk1699 · 2026-08-19
• [source] Ant Group opensources Ling-3.0 Base models with training checkpoints — 智东西 · 2026-08-20

2 near-duplicate retellings: Asleep-Pilot-4142 · niacolhealth