---
title: "Open Weights Take 71.7% of Tokens. Closed Models Keep the Rest."
date: 2026-11-22
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-open-weights-token-share-1122
language: en
---

# Open Weights Take 71.7% of Tokens. Closed Models Keep the Rest.

> A year ago, closed models carried 73.2% of the tokens on one routing platform. Open-weight models now carry 71.7%.

A year ago, closed models carried 73.2% of the tokens on one routing platform. Open-weight models now carry 71.7%.

The pool barely changed size. The occupants swapped.

## Three points make an arch

The numbers come from OpenRouter, cited in a Hugging Face announcement on November 16. In October 2025, open-weight models were 26.8% of tokens. In July 2026 they peaked at 77.5%. At the end of September 2026 they were 71.7%.

Eleven months, three points, one shape.

Two points would have drawn a straight line, and the line would have said open weights only go up. The middle point says otherwise. The series rose, topped out, and gave back 5.8 points in two months.

## The 1.5-point handoff

Percentages split a pool in two, so the other side is easy to compute. At 26.8%, closed models held 73.2%. At 71.7%, they hold 28.3%.

Put the numbers side by side. Closed models had 73.2% in October 2025. Closed models have 28.3% now. Open weights had 26.8% a year ago and 71.7% now.

The share closed models held a year ago sits within 1.5 points of the share open weights hold today. The chair changed. The chair's dimensions did not.

Whether the pool itself grew is not in this data. A share says nothing about volume, and a platform can host a steep mix on a small base.

## The climb was twice as fast as the fall

From October 2025 to July 2026, open weights gained 50.7 points in nine months, about 5.63 points a month.

From July to the end of September, they lost 5.8 points in two months, about 2.9 a month.

The retreat is real, and it is slower than the advance by roughly half. In relative terms, 5.8 divided by 77.5 is 7.5%. Under 6 points absolute, under 8% relative, both readings say the same thing: a pullback, not a reversal.

It also gives the series a ceiling that is only a ceiling among known points. July may not have been the high-water mark. The announcement has no data for the months on either side, so 77.5 is the maximum of three observations, not a proven peak.

## The denominator is tokens

71.7% is a share of tokens routed by OpenRouter. It is not a share of calls, users, revenue or model count.

Tokens measure volume, and volume is sensitive to length. One long request can outweigh many short ones. A user base that asks longer questions will push the token share up without a single new user arriving. The share can move while the people do not.

Revenue is a different denominator, because revenue is tokens multiplied by price and prices differ. Token share does not convert into revenue share, and this data contains no prices. Calls are a third denominator: one long request and one short question count the same as calls and differ enormously as tokens.

One number answers one question.

## One platform&#x27;s ledger

The scope is a single router. 71.7% describes the token mix on OpenRouter. It says nothing about the global token mix, global revenue or global users.

That scope is easy to lose. Add "global" and "market" to the sentence and the claim grows by an order of magnitude without a single digit changing.

There is also a timing gap. The data ends in September. Hugging Face moved its language-model traffic to OpenRouter's routing on November 16. Whatever that traffic did afterward is not in the 71.7%.

## What a router actually does

A router holds a request, picks a model, and returns the answer. Which model it picks is set by rules the platform owns, and the announcement does not describe those rules.

What can be said is positional. The request reaches the router before it reaches a model, and whoever holds the request decides which path the tokens take.

Hugging Face hosting models while handing routing to OpenRouter puts the decision in someone else's hands. The tokens are then counted on the router's ledger, not the host's. That is how a number like 71.7% gets produced. A router does not make tokens. It decides whose tokens they become.

## One sentence of context

Nvidia's interim CEO, in the first earnings call after Jensen Huang's departure, restated the $12.9 billion acquisition of Hugging Face and the company's open-model strategy.

That is background here, not evidence for the share.

## A seat, not a price

A token share is not pricing power. It is the seat from which pricing gets discussed.

With seven of every ten tokens on one router flowing to open weights, the reference quote has a reason to sit on the open side. How far it has moved, and at what price, is not in this data. There are no prices here, only volume.

The composition changed. The entrance moved from the model vendor to the router, and the mix moved from closed to open weights. Whether the price followed is a different ledger, and that ledger has not been published.

Two things are true at once. Open weights took most of the tokens on one platform, and one platform is not the market. The first is a real shift in where volume enters. The second is the limit on what it can be used to prove.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-open-weights-token-share-1122
Free to quote with attribution and a link to the original.
