Ling-3.0-Flash Released

InclusionAI, under Ant Group, released the Ling-3.0-Flash weights under the MIT License.

• Total parameters are 124B, with approximately 5.1B parameters activated per token.
• It activates 8 out of 512 routing experts and 1 shared expert.
• It uses native hybrid linear attention, interleaving 35 KDA layers and 7 gated MLA layers in a 5:1 ratio.
• The research team claims benchmark performance is comparable to or exceeds their own 1T parameter Ring-2.6-1T. However, all figures are self-reported by the development team.

At launch, GGUF and llama.cpp are not supported, and official serving examples require their own SGLang and vLLM forks and 4 GPUs. Although the active parameters are small, the full 124B weights must be held in memory, creating a distinction between inference costs and model hosting costs.