Architecture shortlist

MoE deployment shortlist

Compare open-weight MoE candidates using total and active parameter scope, with clear notes on sparse computation, full weight storage, and distributed serving.

What is included
Seven selected MoE profiles with total and active parameter figures recorded in the directory. The counts have different publisher boundaries and are not normalized compute or memory measurements.
How this list is ordered
Alphabetical by model name. Active-to-total ratios are not used as an efficiency ranking.

Publisher sources are linked for every model. No independent performance scores have been measured. On smaller screens, scroll the table horizontally.

MoE deployment shortlist. Alphabetical by model name. Active-to-total ratios are not used as an efficiency ranking.
ModelTotal / activeContextDeployment note
GoogleSource
~26B / 3.8B256KThe language backbone and vision encoder have distinct parameter scope.Compare model
XiaomiSource
309B+ / 15B1MThe backbone count excludes additional multimodal and prediction modules.Compare model
MiniMaxSource
~428B / 23B1MSparse attention and expert routing both need compatible runtime support.Compare model
Mistral AISource
119B / 6.5B256KUse the total checkpoint scope, not the Small name or active count, for sizing.Compare model
NVIDIASource
30B / 3B1MThe listed checkpoint uses NVFP4; stored and compute precision can differ.Compare model
AlibabaSource
80B / 3B256K3B active does not turn the full 80B checkpoint into a 3B memory budget.Compare model
AlibabaSource
125B+ / 6B256K nativeAdditional n-gram embeddings and the prediction head matter for total storage.Compare model

A sparse compute count is not a resident weight count

Experts that are unused for one token can still occupy memory in a fully resident deployment. Total and active parameters belong in separate columns. Offloading changes where weights are held and how they are transferred, so it should have its own measured configuration rather than borrowing a fully resident result.

Do not rank efficiency from a ratio

An active-to-total ratio omits architectural work, memory traffic, communication, and runtime implementation. Two models with similar ratios may have very different serving requirements. This page intentionally exposes the source numbers without converting them into a quality or speed score.

Measure the system you plan to rent

For distributed configurations, record GPU count, interconnect, parallelism, precision, and request load. Test whether your latency target holds while the service processes realistic inputs. Until those runs are complete, the table is a deployment planning aid and not a recommendation for a particular instance type.

Before you choose hardware

MoE GPU requirements: active parameters vs. total weights explains the next planning step. For the distinction between publisher specifications, calculations, and measurements, read the BenchGrid methodology.