---
title: "100 Billion Downloads: Open Weights Moved From Option to Default"
date: 2026-11-22
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-china-open-weights-downloads-1122
language: en
---

# 100 Billion Downloads: Open Weights Moved From Option to Default

> 10 billion downloads is not 10 billion users. The Ministry of Industry and Information Technology counts cumulative pulls of domestic open-source large models, and that total has passed 10 billion.

10 billion downloads is not 10 billion users. The Ministry of Industry and Information Technology counts cumulative pulls of domestic open-source large models, and that total has passed 10 billion. A pull is a fetch. One person downloading the same model 10 times counts 10 times. A team that pulls the weights once onto an internal network and serves 20 colleagues may count once.

It is a distribution metric, and only that.

Distribution answers one question: how many people took the thing away. It says nothing about whether they used it, for how long, or whether anyone paid. 10 billion pulls can prove the on-ramp is wide. It cannot prove that anyone stayed. Keeping those two questions apart is the only way to see where open weights actually sit.

## The largest open model, in full

Moonshot published the complete weights of Kimi K3.1 Pro, described as the largest open-source model by parameter count. The word complete is the one to stop on. The common practice is to release a smaller sibling, an older version, or inference code without weights. Releasing full weights hands over the most expensive asset the company owns.

You cannot take that back.

Weights are the result of training; code is the method. Publishing the result lets others start from your finish line. Publishing the method lets them see how you got there. Both are called open source, but they hand over different things. The first asks for compute. The second asks for patience.

## An attention curve with no multiplier

Kimi K3 uses an in-house hybrid linear attention mechanism. Yang Zhilin, founder and chief executive of Moonshot Kimi, said the design makes very long text cheaper to compute and very large models more stable to train. Both are comparatives, with no multiplier given.

A direction, not a measurement.

Attention cost is the real budget line for long-context models. Double the length and full attention more than doubles the number of positions that must be compared against one another. Linear attention tries to flatten that curve: let length keep growing while cost grows more slowly. How much flatter is precisely the number the quote does not give, and producing one will take a third-party rerun.

## The cost step DeepSeek left behind

One step back is DeepSeek. Its models from early last year reworked mixture-of-experts architecture and pushed down inference cost and compute requirements at the same time, after which MoE became a direction the whole field explores. Note the subject of that sentence. Not a company, an architecture.

A discount lasts only as long as someone keeps paying for it. Compute saved by architecture travels: it survives a change of country, a change of team, a change of business model. That is why a download matters beyond the file transfer. What arrives with the weights is a cost structure someone else already validated.

## A display maker&#x27;s model

BOE describes its Blue Whale display model as built on a domestic open-source base and covering R&D design, manufacturing, and quality management. Jiang Xingqun is BOE's senior vice president and co-chief technology officer. A panel maker needs yield, process parameters, and defect images. None of that appears on a public leaderboard. It appears on its own lines.

The factory had the data. What it lacked was a base that could read it.

Training one from scratch is expensive and slow. Starting from open weights and continuing the training changes the starting point. This is the most overlooked category of open-weight user: not in the internet industry, not publishing models, not entering evaluations, just wiring a model into a workshop. Their downloads never become a rank on a leaderboard. They become parameters on a production line. How many of the 10 billion pulls fall into this category has not been broken out.

## A base model chosen abroad

Farther out, Singapore's Sea Lion V4 multilingual model uses Alibaba's Qwen open-source model as its base. A multilingual project picked weights released by a Chinese team as its starting point. The weight of that fact is not in the export. It is in the choosing. Whoever picks a base compares licenses, multilingual coverage and maintenance commitments, and then answers to the people who will depend on it.

Downloading is taking. Choosing is staying.

Taking can be reversed. Staying means changing an architecture, a data pipeline, an evaluation set.

## What makes a default

3 users, 3 directions. Moonshot releases full weights, which is supply. BOE builds a production-line model on an open base, which is domestic demand. A Singapore team builds a multilingual model on Qwen, which is demand from abroad. The case for open weights as the default does not rest on the 10 billion. It rests on these 3 things happening at once.

What the 3 share is that none of them trains from zero. The model maker releases weights because it has finished training. The panel maker uses weights because it does not want to start. The Singapore team uses weights because it needs multilingual coverage, not a corpus assembled from scratch. Put the 3 refusals together and the shape of a default appears: training concentrates in a few hands, use spreads across many.

Huang Qi, party secretary of the Institute of Scientific and Technical Information of China, said a group of globally leading open models keeps emerging, with downloads and derivative model counts in the lead, driving the global open-source AI ecosystem. The first half of that sentence is countable. The second half is a judgment others have to accept. Counting is not the same as being ahead.

One line in the argument can be checked directly. The base model field for Sea Lion V4 says Qwen. It is public, and anyone can look it up.

*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-china-open-weights-downloads-1122
Free to quote with attribution and a link to the original.
