China AI Hub China AI Hub

Models / Mixture-of-Experts (MoE) Chinese AI models

Mixture-of-Experts (MoE) Chinese AI models

Mixture-of-Experts models route each token through a small subset of a much larger parameter set, cutting per-token compute cost. This hub lists models documented as MoE — either by an explicit MoE architecture label or by a published active-parameter count (the sparse-activation signature only MoE models publish).

12
models match this dimension
Counted automatically from the models database: architecture contains "MoE" or "Mixture-of-Experts" (case-insensitive) OR parameter_information.active_parameters is present

Active parameters, not total, drive inference cost. Activation ratios from ~1.4% (DeepSeek-V4.1-Flash) to ~5.6% (GLM-5.3-Flash) are a leading indicator of long-run price floors.

Editorial position

Our read: parameter counts are the most misleading number in Chinese AI marketing. A 744B model with 40B active parameters costs roughly like a 40B model to serve and like a 744B model to host. Buyers comparing on size are comparing the wrong figure; activation ratio and memory footprint decide the bill.

Pattern in the data

Distribution of this dimension across matching models
Models matching12
Open weights11 of 12
Providers represented6
Context window range256K – 1024K
Input price range /1M$0.15 – $3
Licence familiesMIT, Apache-2.0, Modified, Kimi, MiniMax, Qwen3.8-Max, proprietary

The data shows activation ratios from roughly 1.4% to 4.3% of total parameters, and that ratio, not the headline parameter count, tracks the price of serving. Two models with the same total size but different activation ratios are not comparable on cost, even when their benchmark scores are.

Matching models

GLM-5.2

zhipu-ai · deprecated

  • Context: 1,048,576 tokens
  • Open weight · MIT (pure open, no regional limits per official model card)

GLM-5.3

zhipu-ai · active

  • Context: 1,048,576 tokens
  • Open weight · Apache-2.0 (per GitHub repo metadata; README has no separate weights-license section - verify per-model HF cards before reuse)

GLM-5.3-Flash

zhipu-ai · active

  • Context: 1,048,576 tokens
  • Open weight · Apache-2.0 (per GitHub repo metadata; README has no separate weights-license section - verify per-model HF cards before reuse)

Kimi K3

moonshot-ai · active

  • Context: 1,048,576 tokens
  • Open weight · Kimi K3 License (permissive MIT-style, but Model-as-a-Service operators with >$20M aggregate revenue over any 12 months must sign a separate agreement; products with >100M MAU or >$20M monthly revenue must display 'Kimi K3' in the UI)

MiniMax-M3

minimax · active

  • Context: 1,048,576 tokens
  • Open weight · MiniMax Community License (custom): free for non-commercial use; commercial use requires prominent attribution 'Built with MiniMax M3' plus written authorization from MiniMax if yearly revenue exceeds US$20M (otherwise a one-time notice to [email protected])

Qwen3.8-2.4T-A95B

alibaba-cloud · active

  • Context: 262,144 tokens
  • Open weight · Qwen3.8-Max License (custom MIT-style: unrestricted use/copy/modify/sell, but products with >100M MAU or >US$20M/month revenue must display the model name; Model-as-a-Service or AI Work Assistant businesses with >US$50M/12-month revenue need a separate license from Qwen)

Original research

Technology

This page is generated from the models database; the member list and count are derived from each model's frontmatter fields, not written by hand. See the full model database and the methodology.