MarkTechPost Compares Token Routing Across 10 Open MoE Models, Including DeepSeek, Qwen, Mixtral, Llama 4 and gpt-oss
MarkTechPost has published a config-verified comparison of how 10 open mixture-of-experts models route tokens. The title names DeepSeek, Qwen, Mixtral, Llama 4 and gpt-oss among the models compared.
MarkTechPost published a config-verified comparison on November 8, 2026 of how 10 open mixture-of-experts (MoE) models route tokens. The article, titled "Which Routing Does Each MoE Model Use? DeepSeek, Qwen, Mixtral, Llama 4, gpt-oss Compared," names those five model families in its headline.
What the comparison covers
The comparison is organized around the following routing and architecture attributes:
- Total and active parameters
- Expert counts
- Top-k (the number of experts selected per token)
- Shared experts
- Gating functions
- Load-balancing methods
According to MarkTechPost, the comparison is verified against model configuration files rather than drawn from marketing descriptions.
Models named
The title lists DeepSeek, Qwen, Mixtral, Llama 4 and gpt-oss. These span several developers: DeepSeek, Alibaba Qwen, Mistral AI, Meta and OpenAI. The remaining models in the set of 10 are not listed in the title.
What this means
MoE models activate only a subset of their parameters for each token, so the routing design shapes how total and active parameter counts relate. The attributes covered here (top-k, shared experts, gating and load balancing) are the ones that define that design. A side-by-side view drawn from config files lets engineers compare these choices across open models using one consistent set of fields.
The full article is available on MarkTechPost: https://www.marktechpost.com/2026/11/08/moe-model-routing-comparison/
Related Articles
Hirundo and Lazarus AI edit weights of Chinese open models to strip censorship for Western clients
Startups Hirundo and Lazarus AI are editing the weights of Chinese open-weight models, including Alibaba's Qwen, to remove state-aligned censorship and harden security for Western enterprises, Rest of World reports. Some researchers question the premise, and some users see no need for it.
Chinese AI labs aim self-improvement at training infrastructure, not model weights, Leiphone survey finds
A Leiphone survey, reported by Pandaily, finds Chinese model developers pursuing recursive self-improvement by pointing AI at training pipelines, inference systems, GPU kernels, sandboxes and agent harnesses. No lab is letting AI rewrite model weights. Zhipu, DeepSeek, ModelBest, Moonshot AI, Alibaba, ByteDance and Xiaomi each disclosed a different piece of the approach.
DeepSeek Harness Reaches 1.54M Daily Users, Plans Official Plugin Marketplace
DeepSeek Harness, the company's open-source agent runtime, had 1.54 million daily active users as of November 4 and consumes about 92 trillion tokens a day. The team plans an official plugin marketplace and a support program for plugin authors, with no launch dates given.
DeepSeek adds "creator mode" to DeepSeek Harness for building custom plugins
DeepSeek has released a "creator mode" that lets users of DeepSeek Harness create custom plugins, including a 3D model renderer and a video editing workbench, according to Bloomberg. The option is designed to help users handle a range of complex tasks.
Comments
Loading...