Open Weight Models in 2026: Qwen and Chinese Labs

Chinese AI models set the size ceiling for open weights

Parameter count is the number of learned values inside a model, and it largely determines how much memory a model needs to run. On that measure, Chinese labs set the ceiling for open weights in 2026.

MiniMax AI and Moonshot AI publish almost nothing below 70B

Across 2026, the biggest and strongest open model from a Chinese lab was larger than any American release in almost every month. China's monthly ceiling ranged from 754B to 2.78 trillion parameters.

In five of the seven months, no American release reached 130B. Two things broke that pattern: NVIDIA shipped Nemotron 3 Ultra in May and June, and Thinking Machines Lab released Inkling.

Several Chinese labs also publish only at the top end. Moonshot, MiniMax, Xiaomi and Z.ai release almost nothing below 70B, so the first model a developer tries from one of them is too large for any machine they own. Xiaomi, Ant Group and Meituan all passed a trillion parameters this year.

The American side is not empty above 100B, but part of it leans on Chinese work. Hugging Face notes that some US releases in that range are built on Chinese models or reuse their artifacts, and gives Inkling (952B) as an example. For original American work, its examples include Arcee AI's Trinity-Large (399B), NVIDIA's Nemotron 3 Super (124B) and Nemotron 3 Ultra (561B).

Alibaba Qwen models and Tencent cover every size from under 1B

Tencent and Alibaba's Qwen team take the opposite approach to Moonshot, MiniMax, Xiaomi and Z.ai. They release models at every size, from under 1B upward. That means a developer can begin with a version small enough to run on their own machine.

If you are choosing a model family, the lowest size in a lab's list tells you whether you can run anything from it on your own hardware. You can check this yourself: open a lab's organization page on the Hub, list its repositories and read the parameter count on each model card. If a lab's smallest model is 70B, plan on reaching it through an API or a rented cluster.

NVIDIA Nemotron and AMD publish the most open source AI models

By number of new model repositories, the top two publishers are the chip makers. AMD and NVIDIA each released more than 200 new repositories, well ahead of every other publisher. LiquidAI is third at around 100.

Hugging Face's explanation is commercial: a freely available model tuned for your hardware is the plainest proof that the hardware works. Hugging Face also says AMD contributed many conversions that let trillion-parameter models run efficiently on US hardware, which it calls a distribution and optimization layer. A repository count therefore measures how much a company publishes, not how capable what it publishes is.

Hugging Face models: open source LLM ranks by downloads and likes

The Hub counts both downloads and likes. The two rankings diverge: Kimi K3 leads on likes, while models under 1B take 83% of all-time downloads.

Kimi K3 earns likes, while small language models earn the downloads

Take the 25 most downloaded repositories of the year and the 25 most liked. Exactly one repository is on both lists.

Not one model published in 2026 makes the download list, and thirteen of its 25 entries date from 2022. The report counts downloads inside the year rather than lifetime totals, so older models get no credit for having existed longer.

The all-MiniLM-L6-v2 model was downloaded 1.55 billion times in seven months and has 5,156 likes. Kimi-K3 was downloaded about 60 times for each like it received.

Hugging Face reads the gap this way. A like marks a release that matters and goes to frontier models in the weeks after launch.

A download marks something wired into a scheduled pipeline, and download counts build up on small, stable models over years. In short, likes show what excites people and downloads show what they already rely on.

Small language models under 1B take 83% of open source downloads

Models under 1B parameters take 83% of all-time downloads, while those above 100B take 1%. The shares are calculated across models that state a parameter count.

Restricting the count to 2026 gives the same picture, with 3% going to models above 70B. Hugging Face's explanation is simple: only small models fit the machines most developers actually have.

The same split shows up publisher by publisher. The table gives the share of each publisher's 2026 downloads that went to models above 70B. Google, Microsoft and IBM Granite are essentially at zero on that measure.

The totals show the same split. Moonshot's frontier-only portfolio drew 37M downloads across the year, while Qwen's spread of sizes reached 2,045M (counting repositories with a declared parameter count), about 55 times more.

Qwen models are the open source LLM base with the most derivatives

A derivative is a model someone built from an existing one, for example by fine-tuning or quantizing it. By that count, Qwen is the base with the most derivatives on the Hub.

Unsloth and Google trail Qwen on Hugging Face models

Qwen-based models account for 151,448 derivatives on the Hub, well ahead of Google in second place with 82,506. Measured against Meta, Qwen's count is 2.6× everything Meta has on the Hub and 4.7× what the Llama repositories hold.

The third-largest source is Unsloth, a community account that publishes quantized and fine-tuning-ready builds, and many of them are Qwen builds. Qwen derivatives have been growing by roughly 180–210 new repositories per day across the first seven months of 2026.

Hugging Face gives three reasons: a steady release schedule that keeps refreshing the whole family, coverage of many sizes and uses, and Apache 2.0 licensing. The size range lets a developer stay in one family from a small local model to a larger deployment, and the license lets anyone modify, redistribute and commercially use the results.

Most of that lead comes from other developers, not from Qwen. Qwen itself published only 54 of the 28,531 GGUF conversions of its models on the Hub.

Llama vs Qwen: 39.6 million GGUF downloads a month against 7.5 million

GGUF is a file format for running a quantized model on a local machine. Qwen gets 39.6 million GGUF downloads a month, against 7.5 million for Llama. Gemma sits at 20.8 million, so Qwen has close to twice Gemma's volume and over five times Llama's.

Llama is not short of supply. Llama-derived GGUF repositories slightly outnumber Qwen's, so the supply is comparable, yet Llama's monthly downloads are about a fifth of Qwen's. The gap does not come from how much is on offer.

llama.cpp raises the local LLM ceiling to mixture of experts models

A mixture of experts model splits its parameters into many specialist blocks and uses only some of them for each token. The open question is how anyone outside a data center runs one. According to the report, the ceiling for local models moved with llama.cpp.

ggml joined Hugging Face, and DeepSeek models have GGUF builds

In February the ggml team joined Hugging Face, and the July snapshot of the Hub carries GGUF builds of DeepSeek-V4-Flash (roughly 284B parameters) and Kimi-K3 (about 2.8 trillion). The ggml project stays fully open source and community-governed, and Hugging Face calls it the most important project in local inference. For Hugging Face, local inference now means a trillion-parameter mixture of experts split across a few consumer machines.

Library tags point the same way. The table shows how fast repositories declaring each library grew in 2026.

The report's reading is that the layer deciding where a model can physically run is growing several times faster than the modeling core. That layer includes local inference formats, Apple silicon and robot control stacks.

Labs publish few official GGUF builds, so community builds fill in

Across the ten largest model families, the labs publish very few official GGUF conversions, yet GGUF files are often the versions developers actually use. The files in use are therefore often community builds. Hugging Face suggests labs ship an official conversion at release, document their quantization choices and sign the files, or else work with existing contributors such as Unsloth.

Before you download a GGUF file, check who published it on the model page. For the ten largest families the answer will often be a community account rather than the lab, and that conversion is not necessarily the build the lab tested.

Apache 2.0 covers most open weight vs open source releases from China

An open weight model has published weights you can download and run under a license. Open source in the stricter sense also expects the training code and data to be public. For Chinese releases above 20B, the license on the model page is usually Apache 2.0 or MIT.

Chinese LLM releases above 20B: 59% Apache 2.0 license, 22% MIT

59% of the Chinese releases above 20B in 2026 carry Apache 2.0, and another 22% carry MIT. The sample is 178 releases, and almost none of them have non-commercial restrictions.

Two of the newest releases are the exception. In the last few weeks before the report, Kimi K3 and Qwen 3.8 2.4T started adding some non-commercial restrictions and revenue share requirements to their licenses. Check the current license field on the model page for any recent large release.

DeepSeek models ship under MIT, while US open weight LLM terms vary

DeepSeek and Z.ai release models from 700 billion to 1.65 trillion parameters under plain MIT, while American terms in the same size band vary. The table splits the American releases in that band by license.

Hugging Face reads these releases as a bet on something other than license fees. The payoff has to come from API and cloud sales, hardware and platform standing, or simply a place in the ecosystem.

Hub counts cannot name the best open weight models

The report closes with limits on how far its own numbers go. Hub numbers leave out closed APIs and private deployments, so they describe community activity rather than market share.

Hub activity is not market share for Chinese open source models

Hugging Face says none of these counts is a direct measure of quality, commercial adoption or market share. Downloads, likes, derivatives and releases each capture a different part of ecosystem activity. Likes reflect community attention, and derivative counts show how much developers build on an existing model.

The analysis covers Hub activity in the first seven months of 2026. The data is also heavily skewed. Public model repositories rose from 2.43 to 2.96 million over the period, while 1.5% of them account for 99.2% of every download, and roughly 85.6% have fewer than 200 lifetime pulls.

Closed APIs and private deployments sit outside the Hub numbers

Downloads count use inside the Hub only, so API use, private deployments and other distribution channels are missing. A company that serves a model to its own users, or reaches one through a cloud provider's API, adds nothing to these counts beyond the download that fetched the weights, if there was one.

What the data can tell you is what the community publishes, what it depends on and what it builds on. On those questions the answers are Chinese labs at the top of the size range, small models in the download tables and Qwen as the starting point for derivatives. Which model is best for your task is a question for your own test set.