Ling 2.0#

Ling 2.0, documented in this repository under the Bailing family, is a high-sparsity Mixture-of-Experts language model family from inclusionAI. Megatron Bridge supports the Ling 2.0 MoE checkpoints through the Bailing bridge with auto-detected Hugging Face configuration and weight mapping.

Supported Variants#

Variant

Hugging Face ID

Notes

Ling-flash-2.0

inclusionAI/Ling-flash-2.0

100B total, 6.1B active

Ling-flash-base-2.0

inclusionAI/Ling-flash-base-2.0

Base checkpoint, 100B total, 6.1B active

Ling-mini-2.0

inclusionAI/Ling-mini-2.0

16B total, 1.5B active

Ling-mini-base-2.0

inclusionAI/Ling-mini-base-2.0

Base checkpoint, 16B total, 1.5B active

Architecture Notes#

  • High-sparsity MoE with 256 routed experts and top-8 routing.

  • Sigmoid routing with QK-Norm and Half RoPE.

  • Custom Hugging Face model code is required, so conversion and inference commands use --trust-remote-code.

Examples#

For checkpoint import/export, round-trip validation, and inference commands, see the Bailing examples README.