An MoE model name can highlight a small active-parameter number while the checkpoint contains a much larger collection of experts. That distinction matters immediately for a GPU quote. Active parameters help describe a token’s computation path; they do not establish how much model storage must be available or how experts are placed across a serving group.
1. Read both parameter counts
Qwen’s official Qwen3-235B-A22B model card lists 235 billion total parameters and 22 billion activated parameters [1]. Treat those as different fields in the specification. Ask the supplier which checkpoint is being offered, including quantization and revision. Neither the active count nor the filename tells you the exact allocated bytes, because tensor formats and packaging still need inspection.
2. Start the weight screen with total stored parameters
In a deliberately simplified scenario where all 235 billion parameters use two bytes, the raw payload is 470 GB, about 437.7 GiB. Multiplying 22 billion by two gives 44 GB, but that is not the complete weight payload of this model. This arithmetic is not a recommendation for a particular GPU count. Real checkpoints may mix precisions and add scales or other data.
| Specification field | Useful for | Does not prove |
|---|---|---|
| Total parameters | Initial weight-capacity screen | Exact runtime allocation |
| Active parameters | Understanding selected computation | Entire model fits in that payload |
| Expert placement | Per-device capacity and traffic | Balanced load for every request |
| Cache policy | Live context capacity | Weight residency |
3. Ask where every expert resides
A fully resident deployment must account for all required weights across the serving group. An offloaded design changes that placement but introduces another resource path; it does not erase the experts. Request a diagram showing shared layers, expert shards, any replicated components and host-resident data. vLLM documents distinct parallel approaches for MoE layers [2]. Require the actual supported arrangement for the chosen engine version.
4. Test routing-sensitive traffic
A generic prompt set may not exercise the same expert traffic as your domain. Build acceptance inputs from intended languages, document types and tool-use patterns, while protecting confidential data. Compare per-device utilization, communication and latency across that mix. The practical question is whether one part of the group becomes the limiting resource under realistic routing, not whether average utilization appears attractive in a screenshot.
5. Keep KV cache in a separate budget
Sparse expert activation does not remove the attention cache needed by the model’s generation path. Use the actual attention architecture, retained context and concurrent sequences to calculate a cache scenario. Then measure the complete engine. Weight quantization, cache quantization and host offload are separate configuration decisions; ask which one produced any advertised capacity improvement and whether quality or latency changed.
6. Specify an acceptance boundary
Compare supplier offers at the same model revision, precision, topology and service target. Request both warm steady-state and cold-start behavior if offload or cache reuse is involved. Include a failure or restart exercise for the serving group and identify the unit of replacement in the support plan. A model distributed across several cards creates a different operational dependency from independent small-model replicas.
- List total and active parameters in separate RFQ fields.
- Require per-GPU weight and cache allocations.
- Reject capacity claims calculated from active parameters alone.
Discuss your configuration
Size MoE weight capacity from the complete representation and expert placement; assess active computation and communication separately.
Explore the hardware catalog ↗Discuss your configuration ↗Sources & methodology
Parameter counts follow the official model card [1]; MoE parallel-serving options follow [2]. The 470 GB example is a homogeneous two-byte storage assumption, not a measured checkpoint or runtime footprint. Sources verified 24 October 2026. Estimates and purchasing recommendations are GPUSupplier editorial analysis, not our laboratory measurements.