BWhat changed
Integrated reassignment of capacity-dropped tokens, expert-specific learning-rate scaling and combined head/layer KV sharing into an open MoE training system, with published recipes and model checkpoints.
Scientific scope
Integrated reassignment of capacity-dropped tokens, expert-specific learning-rate scaling and combined head/layer KV sharing into an open MoE training system, with published recipes and model checkpoints. Mixture-of-experts, shared and routed experts, grouped-query attention and cross-layer attention have predecessors. This contribution combines recycle routing and expert-aware training; its scaling fits and performance comparisons remain specific to the reported experiments.
Evidence
The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Mixture-of-experts, shared and routed experts, grouped-query attention and cross-layer attention have predecessors. This contribution combines recycle routing and expert-aware training; its scaling fits and performance comparisons remain specific to the reported experiments.
Contributors
- Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.
- Hunyuan-Large: An Open-Source MoE Model with 52 Billion Activated Parameters by Tencent
- Tencent Hunyuan Team: recycle routing and expert-aware MoE training — official research provenance
Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction