Research

MoE Active vs Total Parameters: A Hardware Guide

Why sparse mixture-of-experts models need memory for every parameter but only compute for a few. Active versus total parameter counts for the 2026 open-weight models, and how to size hardware correctly.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: July 2026

The most common and most expensive mistake in local AI hardware buying in 2026 comes from one misunderstanding: people size their machine against a model's active parameter count when they should size it against the total. The two numbers can differ by a factor of thirty.

The Rule

Memory footprint tracks total parameters. Speed tracks active parameters.

A sparse mixture-of-experts model contains many expert subnetworks and routes each token to a small subset. Only the selected experts do arithmetic, which is why these models are fast. But the router can select any expert for any token, so every expert must be resident in memory and ready. You pay for all of them in capacity and only some of them in compute.

The 2026 Open-Weight Models

ModelTotal parametersActive per tokenSparsity ratioApprox. memory at 4-bit
Kimi K32.8T~16 of 896 expertsVery high~1,400 GB
GLM-5.2753B~40B~19x~377 GB
MiniMax M3428B~23B~19x~214 GB
Qwen3-Coder-Next80B~3B~27x~40 GB
Nemotron TwoTower~60B~3B per tower~10x~30 GB

Qwen3-Coder-Next is the instructive case. Its 3B active parameters give it the decode speed of a very small model, and reviewers report throughput comparable to models with 10 to 20 times more active parameters. But its 80B total means roughly 40GB of memory at 4-bit, which is more than an RTX 5090 has. It is simultaneously fast enough for a laptop and too large for a high-end consumer GPU.

What This Means When Buying

Look up the total parameter count, divide by two for a 4-bit estimate in gigabytes, add 20 to 30 percent for key-value cache and runtime overhead, and check that number against the memory your candidate hardware exposes to the model. Then, separately, look at active parameters to predict speed. Doing only the second step is how people end up with a fast GPU that cannot load the model they bought it for.

If you haveLargest total parameters at 4-bitExample models that fit
16GB VRAM~24BDense models to ~24B, small MoE
32GB VRAM (RTX 5090)~50BNemotron TwoTower, dense 32B
64GB unified~100BQwen3-Coder-Next, gpt-oss 120B tight
128GB unified~220B120B-class comfortably
640GB (8x H100)~1.2TGLM-5.2, MiniMax M3

Why Labs Build This Way

Sparsity decouples capability from inference cost. A 753B model with 40B active has much of the knowledge capacity of a very large dense model at roughly the serving cost of a 40B one. For a lab paying for inference at scale that trade is overwhelmingly favourable, and it is why nearly every significant open-weight release in 2026 is sparse. The cost is pushed onto whoever provides the memory, which at a hyperscaler is cheap and at a workstation is the binding constraint.

The counter-movement is models such as Qwen3-Coder-Next that deliberately keep total parameters modest so the whole thing fits on hardware people own. See MoE adoption across open weights.

Brand Visibility Implications

Sparse architectures are why frontier-quality models now run at small-model speed, and speed is what made local coding agents and private enterprise deployments practical. Every improvement in the sparsity ratio moves more brand-relevant inference off observable endpoints. See the local-LLM visibility blind spot.

Methodology

Vendor specifications come from NVIDIA, Apple, and model-card publications. Throughput figures aggregate community benchmark reporting from the llama.cpp discussions, the MLX repository, and published independent test suites. Single-stream decode unless stated otherwise. Ranges rather than point values are used wherever independent runs disagree, which is most of the time: quantisation format, prompt length, thermal state, and runtime version each move these numbers by more than the differences being measured. Treat every figure as an order-of-magnitude guide, not a specification. Updated quarterly.

How Presenc AI Helps

Presenc AI measures brand representation across the sparse open-weight models that dominate 2026 deployments, hosted or private.

Frequently Asked Questions

All of them. The router can select any expert for any token, so every expert must be resident and ready. Memory footprint tracks total parameters while compute and speed track active parameters. This is the single most misunderstood point in sizing hardware for sparse models.
Take the total parameter count, halve it for a 4-bit estimate in gigabytes, then add 20 to 30 percent for key-value cache and runtime overhead. A 753B model needs roughly 377GB of weights at 4-bit, so budget around 450-490GB in practice.
It activates only about 3B of its 80B parameters per token, giving it the decode speed of a very small model. But all 80B must be resident, which is roughly 40GB at 4-bit, more than an RTX 5090's 32GB. It is fast enough for a laptop and too large for a high-end consumer GPU at the same time.
Sparsity decouples capability from inference cost. A 753B model with 40B active has much of the knowledge capacity of a very large dense model at roughly the serving cost of a 40B one. That trade is strongly favourable for labs serving at scale, and it shifts the cost onto whoever supplies the memory.

Track Your AI Visibility

See how your brand appears across ChatGPT, Claude, Perplexity, and other AI platforms. Start monitoring today.