Ten open mixture-of-experts models compared on how they route tokens
single source· 1 articles · confidence: low · first seen 2026-11-08 10:27 UTC
What this means for you
A reference table, not a result: nothing in it says which routing scheme is better, and no benchmark scores or evaluation dates are attached. It is useful if you are picking between these families for a fine-tune, or want to know how much of each model actually runs per token. Otherwise, nothing to act on.
MarkTechPost has published a side-by-side comparison of how ten open mixture-of-experts models route tokens — the mechanism that sends each token to a few experts rather than the whole network, so only part of the model runs per request. It lists total and active parameters, expert counts, top-k selection (how many experts a token reaches), shared experts, gating functions and load-balancing methods. The post says the details were checked against each model's configuration rather than inferred. DeepSeek, Qwen, Mixtral, Llama 4 and gpt-oss are among those covered; no benchmark scores or evaluation dates are given.
Key facts
- ·The comparison covers 10 open mixture-of-experts models. source
- ·Models covered include DeepSeek, Qwen, Mixtral, Llama 4 and gpt-oss. source
- ·Compared attributes are total and active parameters, expert counts, top-k, shared experts, gating functions and load-balancing methods. source
- ·The post describes the comparison as config-verified, checked against model configuration rather than inferred. source
- ·It was published on 8 November 2026. source
- ·The post reports no benchmark scores or evaluation dates. source
What the sources say
- MarkTechPost — A reference table of routing settings across ten open models, checked against published configs.
Sources
The original reporting. Follow these — they did the work.