Blog · Concepts · September 20, 2026 · 6 min read
Why a “235B” Model Runs Like a 22B: Mixture-of-Experts Explained
.webp)
Model names like Qwen3-235B-A22B look impossible at first: a 235-billion-parameter model that generates as fast as a 22-billion one? The trick is an architecture called Mixture-of-Experts (MoE), and understanding it will change how you read every model name on this site.
How MoE works
In a classic “dense” model, every parameter participates in processing every token. In an MoE model, the feed-forward layers are split into many parallel “experts,” and a small router network sends each token down only a handful of them — usually 2 to 16 out of dozens or hundreds. The result:
- Total parameters (all experts combined) decide how much VRAM you need — every expert must sit in memory even if it is rarely used.
- Active parameters (the few experts per token) decide how fast it generates and how much compute each token burns.
That is what the name encodes: 235B total, Active 22B. A token gets big-model knowledge because the router can pick from 235B parameters’ worth of capability, but only pays the compute of 22B.
Examples you will meet
| Model | Total | Active | What it means in practice |
|---|---|---|---|
| GPT-OSS 20B | 20.9B | 3.6B | Runs briskly on modest GPUs |
| Qwen3 30B A3B | 30.5B | 3.3B | Feels like a small model, knows more |
| Mixtral 8x7B | 46.7B | 12.9B | The MoE that made the pattern mainstream |
| Qwen3 235B A22B | 235B | 22B | Flagship knowledge, workstation-class memory |
The catch: memory still needs the whole model
Speed is small-model, but capacity is big-model. Qwen3-235B-A22B at Q4 needs roughly 120+ GB of VRAM no matter how snappy it feels — the unused experts still have to live somewhere. This is exactly why our compatibility panel sizes models by total parameters, while the speed estimate benefits from the active count for MoE architectures. A 235B-A22B that fits will generate far faster than a dense 70B of similar knowledge depth.
Should you prefer MoE?
If your hardware has abundant memory relative to your compute — Apple Silicon with lots of unified memory, or multi-GPU rigs — MoE models are exceptional value: flagship-level knowledge at mid-range speed. On a single 8–12 GB card, stick to small dense models or small-MoE entries like Qwen3-30B-A3B, which punch far above their active-parameter class. The fit badges will tell you which side of the line you are on.