Las estrategias de enrutamiento MoE determinan qué expertos procesan cada token, lo que impacta directamente en la calidad del modelo, el equilibrio de carga y los costos de servicio en los principales modelos abiertos.
Resumen generado con IA. Consulta el original para comprobar el contexto y la precisión.
Este es el extracto disponible del feed. Puede estar abreviado o en el idioma original.
Every Mixture-of-Experts (MoE) model repeats one decision for every token: which experts should process it. That decision belongs to the router. MoE routing strategies shape quality, load balance and serving cost in every major open model. This guide explains the 7 routing strategies used in production, and the models that use each one. TL;DR Size: Routers choose from 8 experts (Mixtral) to 384 (DeepSeek-V4-Pro), activating 1 to 8 per token. Runs on: Routing cost shows up in multi-GPU expert parallelism, where each choice becomes all-to-all network traffic. Performance: Sigmoid scoring with aux-loss-free bias balancing is the dominant 2026 recipe for frontier open MoE models. Best: Loss-free balancing cut MaxVio from 0.72 to 0.04 at 1B, with lower perplexity. Worst: Expert choice leaks future tokens in causal models; one test showed a roughly 10% abnormal loss drop. Bottom line: Best thing: bias-based balancing adds no interfering gradients. Worst thing: no router removes the need for serving-time load fixes. What is MoE routing? MoE routing is the mechanism that sends each token to a small subset of expert networks. A router scores every expert, keeps the top candidates and weights their outputs. Only the selected experts run. Mixtral 8x7B, for example, gives each token access to 47B parameters but uses 13B active parameters. video created by marktechpost.com The 7 MoE routing strategies used in production 1. Softmax top-k token choice Each token scores all experts with a linear layer and keeps the k highest. A softmax over those k scores sets the mixing weights. Mixtral picks 2 of 8 experts. gpt-oss-120b picks 4 of 128. Qwen3-235B-A22B picks 8 of 128. An auxiliary loss usually keeps expert load even. Used by: Mixtral, gpt-oss, Qwen3. 2. Top-1 switch routing with a capacity factor Switch Transformer showed that routing each token to 1 expert preserves quality and cuts router compute. Each expert gets a fixed buffer: tokens per batch divided by exper
Una ayuda automática para ordenar, basada en fuentes, palabras clave y entidades, en una escala de 1–12. Alta: 9–12; media: 6–8; baja: 1–5. No mide urgencia, verifica noticias ni representa respaldo editorial.
Arete-27B de Nous Research elimina el sesgo alineado con el estado chino de los modelos Qwen, ofreciendo a los desarrolladores una IA de alto rendimiento con pesos abiertos y sin propaganda.
NVIDIA Dory 4B permite que un solo modelo funcione con cuatro presupuestos de cálculo diferentes mediante la repetición de capas intermedias, ofreciendo un rendimiento flexible sin necesidad de volver a entrenar.
Kinference simplifica la gestión del servicio de modelos de IA en Kubernetes al enrutar inteligentemente las solicitudes a los trabajadores de vLLM, optimizando el rendimiento y el uso de recursos para los equipos que implementan múltiples modelos.