Xing4.0 arrives as a 29B MoE with 4B active params

inclusionAI has published Xing4.0-29B-A4B, a text-focused large language model built on a mixture-of-experts (MoE) architecture. The model is available on Hugging Face under a permissive Apache-2.0 license, which allows both commercial and research use.

The headline detail is the balance between total and active parameters. Xing4.0 carries 29 billion parameters overall but activates only about 4 billion per token, following the increasingly popular MoE playbook of routing each input through a small subset of specialized experts.

Why it matters

The appeal of a sparse design like this is efficiency. By keeping the active parameter count near 4B, the model can aim for the quality of a larger network while running with the compute and memory footprint closer to a much smaller one — useful for teams that want capable text generation without heavyweight inference costs.

• 29B total parameters, roughly 4B active per token
• Mixture-of-experts architecture
• Text-only (LLM) modality
• Apache-2.0 license

As an initial release for the Xing family, the model's real-world performance and context handling remain to be seen; inclusionAI has not detailed benchmarks or context length in the release record. Interested developers can find weights and any accompanying documentation on the Hugging Face repository.