Active vs Total Parameters: The Number Spec Sheets Hide

Model announcements lead with the total parameter count because it is the largest number available. It is also the number that tells you least about what the model can do.

For any mixture-of-experts model, the total describes memory and the active count describes computation. Those are different things, they differ by one to two orders of magnitude, and comparisons that use the wrong one produce conclusions that are not merely imprecise but backwards.

What each number means

A dense model runs every parameter on every token. Total and active are the same number, and the spec sheet is honest by construction.

A mixture-of-experts model holds many parameters and routes each token through a small subset. Every expert must be resident in memory, because you cannot predict which the router will pick, but only the selected ones do arithmetic on any given token.

So the total is a hosting cost and the active count is a thinking budget. Kimi K2.6 illustrates the structure clearly: 384 experts, of which 8 are routed and 1 is shared, meaning nine experts process each token and the rest sit in memory waiting. How the routing works covers the mechanism in detail.

The 2026 field, both numbers

Reading the current field with both figures reorders it considerably.

• Kimi K3 — roughly 2.8T total, 104B active. The highest active count in the open-weight field.
• DeepSeek V4 Pro — roughly 1.6T total, 49B active.
• GLM-5.2 — roughly 744B total, around 40B active.
• Kimi K2.6 — roughly 1T total, 32B active, across 384 experts.
• DeepSeek V4 Flash — roughly 13B active.
• Qwen 3.6 27B — dense, so 27B is both numbers.

Sorted by total, K3 leads DeepSeek V4 Pro by about 1.75 to one and Qwen 3.6 27B by about a hundred to one. Sorted by active, K3 leads V4 Pro by about two to one and Qwen by under four to one. The second ordering is the one that predicts capability.

The comparison this fixes

Take the most common bad comparison in the field: a 1.6T mixture-of-experts against a 27B dense model, described as sixty times bigger.

On the figure that governs computation it is 49B against 27B — under two to one. That is why a well-trained dense 27B lands within a few points of a trillion-parameter model on bounded benchmarks. Qwen 3.6 27B reports 77.2 percent on SWE-bench Verified against DeepSeek V4 Pro's 80.6. The gap is real. It is nothing like sixty to one, and it never could have been.

Whenever you see a headline total, find the active count before forming an opinion. If a model card does not publish one, that omission is itself information. What a parameter count actually measures covers why neither number is a capability metric on its own.

Which number governs which cost

Total parameters govern memory, and memory is what determines whether you need one accelerator or a cluster. Kimi K3's weights are roughly 1.6TB — a multi-node deployment with fast interconnect, because tokens routed to experts on other nodes require cross-node communication.

Active parameters govern compute per token, which drives inference latency and, indirectly, the price a provider charges you. It is not an accident that the price ordering across the field roughly tracks the active-parameter ordering: Kimi K3 at $3 input per million, GLM-5.2 around $1.40, DeepSeek V4 Pro around $0.44, V4 Flash around $0.14.

This is precisely why labs build mixture-of-experts models. Total parameters buy knowledge; active parameters cost money. Decoupling them lets a lab add capacity without adding per-token cost, which is an excellent trade for anyone serving millions of requests and a poor one for a single team hosting the model themselves. Dense against mixture-of-experts works through that inversion.

Where active parameters stop predicting

The active count is the best single proxy available, and it is still only a proxy. Two models with identical active counts can differ substantially, because training data quality, training compute, post-training and routing design all matter.

The trend that most complicates the proxy is that training quality has been improving faster than scale. A 27B trained in 2026 is a materially different object from a 27B trained in 2024 at the same architecture and the same count. Comparing active parameters across generations is therefore much less reliable than comparing within one.

Use the active count to sanity-check claims and to reject bad comparisons, not to rank models that are close. For ranking, use benchmarks in the right variant and then your own evaluation. Benchmarking on your own work covers a lightweight version.

Reading a spec sheet in practice

A short procedure covers most cases. Find the active count; if it is absent, assume the model is dense or that the omission was deliberate. Use the total to estimate whether you could host it, at roughly two bytes per parameter at 16-bit precision before cache and overhead. Use the active count to estimate reasoning depth and expected price.

Then check the context window separately, because it is independent of both numbers and is where families surprise people — Kimi K3 ships 1M while K2.6 ships 256K, inside the same family.

The one-line rule: total is memory, active is thinking, and the announcement leads with memory. Small against large models covers what to do once you have both numbers.