Ant Group's inclusionAI Team Releases Ling 3.0 Flash FP8 Under MIT License
Ant Group's inclusionAI team released Ling 3.0 Flash FP8 on August 4, 2026—a 124-billion-parameter Mixture-of-Experts (MoE) language model designed for production-scale developer deployment. Released under the permissive MIT license, the model represents a major step forward in compressing high-capacity reasoning capabilities into a form factor that is cost-effective to run in enterprise environments.
The model uses a native hybrid-linear MoE architecture (part of the BailingMoeV3 family), combining an alternating 5:1 stack of Kimi Delta Attention and gated Multi-head Latent Attention (MLA). By implementing a sparse 1/64 routing mechanism, it only activates 5.1 billion parameters per token. The official FP8 checkpoint utilizes blockwise (128x128) E4M3 expert weights, allowing it to drastically reduce memory usage without sacrificing accuracy on long-context operations. It supports a native 128K context window, which developers can extend up to 256K using YaRN.
Because the release uses a custom bailing_hybrid architecture, running Ling 3.0 Flash FP8 currently requires specialized forks of SGLang or vLLM for inference. The model is designed to replace the previous 1-trillion-parameter Ring-2.6-1T flagship, proving that sparse, quantized architectures can deliver comparable agentic and coding performance at a fraction of the hardware cost. This open release follows a broader industry trend of labs publishing highly quantized checkpoints directly to Hugging Face, enabling startups to bypass closed APIs in favor of self-hosted infrastructure.
