Skip to content

Search across models, tools, services, companies, providers, APIs, plans, articles and news.

How MoE models work — and why their pricing is so confusing

How MoE models work — and why their pricing is so confusing

Before MoE, every answer meant running the whole model. Mixture of experts broke that rule: the network is built from dozens of "experts", and a tiny router wakes only two or three of them per token. The numbers live in the model names themselves: DeepSeek V3 ships 671B parameters with 37B active; Qwen3, 235B with 22B active. A giant model, billed at mid-size weight.

This is where the market\u2019s most confusing part begins: "parameters" no longer equals cost. Two models with 20B active parameters — one may be distilled from a frontier model and run free on your laptop, the other may demand expensive GPUs. The columns you actually buy: active parameters, tokens per second, retrieval quality in your language, and per-token input/output price. The site\u2019s model table lines up exactly these columns.

Practical wrap-up: MoE was the industry\u2019s answer to giant-model cost and is now the default architecture of most Chinese frontier models; whenever a spec table leaves you guessing in the "parameters" column, ask "how much is active?" and "what is the per-token price?" — the rest is marketing noise. The industry moved from "bigger model" to "smarter execution", and MoE is precisely that smarter execution.

Back to blog