Who's winning open-weight AI?
Everyone gives the same one-word answer. I wanted the number behind it, so I stitched together three years of usage data: official records where they exist, the Wayback Machine where they don't.
The question sounds settled. DeepSeek and Qwen headline every discussion of open-weight models; Llama releases land to polite applause. But "China is winning open weights" is a vibe, not a measurement. How much? Since when? Taken from whom? Snapshots of the recent share make headlines every few weeks. The history, and who the share was taken from, nobody had. So I built it.
The usage data comes from OpenRouter, the one place that publishes model-level token counts across hundreds of models. Its official data API only reaches back to January 2025, and this story starts well before that. For the earlier years there's the Internet Archive, which has been quietly photographing the page since 2023, each photo carrying about a week of daily numbers in its source. 119 snapshots cover those early years; OpenRouter's official daily dataset takes over from January 2025. Hugging Face download stats are the cross-check. Highlighted terms are in the glossary; hover or tap.
Each band is a bloc's share of open-weight model traffic on OpenRouter
One chart, four eras
Read the bands left to right and the whole history of open-weight AI plays out.
2023: the hobbyists
Before open models were a geopolitical story they were a hobby. The biggest open-weight model by traffic in late 2023 wasn't from a lab at all. It was MythoMax, a 13-billion-parameter fine-tune someone blended for roleplay chat, with Meta's Llama 2 under the hood. Community creations like it carried 95% of open-weight tokens that November. Worth remembering when the sector gets talked about in billions: this market was born in Discord servers.
Early 2024: Europe's six months
Mixtral changed that. Mistral's December 2023 release was the first open model that felt close to GPT-3.5, and within three months French models carried over half of open-weight traffic. It didn't last, but it was the first time the crown left the community's head.
Late 2024: Llama summer
Llama 3 gave American labs their turn, a bit over half of open-weight tokens at peak. Here's the strange part: through all of 2024, with Meta shipping its best work, the single most-used open model was still MythoMax. Llama's traffic split across dozens of sizes and variants while the roleplay crowd kept hammering one 13B model. Adoption is weirder than press releases suggest.
2025 on: the takeover
Then DeepSeek released R1 in January 2025, and the blue band starts eating the chart. 27% that month. 59% by April. 73% by November. 88% now, with the American share down around 10%, most of it OpenAI's gpt-oss. The throne tells the story even faster than the bands do:
The #1 open-weight model by tokens served, each quarter. Color marks the lab's origin.
Since R1, the crown has only moved between Chinese labs: DeepSeek to MiniMax to Qwen and back to DeepSeek. The competition that used to happen between countries now happens inside one.
Why did it flip this hard? Price is the obvious answer, but the traffic itself changed underneath. Over the same stretch, coding went from roughly 11% of router workload to more than half, per OpenRouter's own study (via Wing's analysis). The market's center of gravity moved from roleplay chat to coding agents, and Chinese labs won the price-performance fight in exactly that workload. So the takeover isn't only models improving. It's the market relocating to the ground where they were strongest.
I didn't tell the pipeline when Mixtral, Llama 3, or R1 shipped. The bands inflect at all three release dates on their own. That's the strongest evidence the reconstruction works.
The part nobody mentions: open weights nearly lost
Winning open weights only matters if open weights matter. For most of 2025, that was in genuine doubt.
The rest goes to closed API models: GPT, Claude, Gemini and company
In early 2024, four of every five tokens on this platform went to models you could download. Then frontier APIs pulled away on capability and took the traffic with them; by spring 2025 open weights were down to a quarter and falling, and "open models are a dead end" was a defensible take. The climb back to 70% is almost entirely blue-band traffic.
China didn't just win the open-weight market. China is the reason there's still an open-weight market to win.
A second opinion: downloads
Token traffic is one router's view of the world, so I checked it against a different population entirely: developers pulling weights from Hugging Face to run themselves. Downloads have their own measurement trap, though. A third of them go to repackaging accounts, the people who quantize models so they run on smaller hardware, and counting those as neutral hides who actually made the thing. So every re-upload gets credited back to the lab that trained its base model. Watch the correction move the numbers:
Share of top-800 text-generation downloads on Hugging Face, trailing 30 days
The unattributed pile shrinks by almost half, and most of it flows to China, because the quantizer community overwhelmingly repackages Qwen and DeepSeek. After the correction: 54% China, 26% United States. Different platform, different metric, different failure modes, same winner. One detail says plenty: the most-downloaded American model is GPT-2, a 2019 research artifact coasting on tutorials. China's is Qwen3, released last year.
Four windows, one question
Any single platform can be dismissed as unrepresentative, so here's the same question asked everywhere I could find an answer: router traffic, developer downloads, the self-hosted crowd pulling models through Ollama, and ModelScope, Alibaba's model hub, which sees the Chinese domestic population Western platforms never will.
Blue is China, orange is the United States, gray is everyone else
They don't all agree, and the disagreement is the finding. Where models run as a business, China dominates: 88% of routed open-weight tokens, 90% of domestic Chinese downloads. Where individuals pull models onto their own machines, America still leads: Llama, Gemma and Phi hold 53% of Ollama's cumulative pulls against China's 30%. Part of that is a real difference in populations. Local models skew small, English-first, and bundled into tutorials. Part is arithmetic: Ollama's counts accumulate since 2023, so they carry two years of Llama-era history the way a career batting average carries your rookie season.
Read together: China won the server. America, for now, holds the laptop. And the throne timeline above suggests which way the laptop is drifting.
Even four windows don't see everything. Enterprises calling OpenAI directly and private clusters running at scale are invisible to every public source, mine included. The first is closed-model territory by definition. The second nobody can measure, and anyone claiming to is guessing.
So, who won?
Chinese labs, in three windows out of four, by margins that make the received wisdom look timid: nearly nine-tenths of open-weight serving, a majority of properly-credited downloads, nine-tenths of the domestic hub, and the entire revival of open weights as a category. The one place America still leads is the laptop, on a cumulative count that owes a lot to history. If current momentum means anything, that lead is living on borrowed time.
The honest asterisk is that tokens aren't dollars. Chinese models win volume partly through aggressive pricing and free tiers, and the closed frontier still collects most of the industry's revenue. Price can buy tokens, though it can't buy downloads, which cost the same for every model and land on the same winner. So the cleaner statement: winning open weights means winning the high-volume, low-margin half of the market, plus whatever it's worth to be the foundation everyone else builds on. That second thing is hard to price and might be the whole game. This page measures usage; it won't settle valuation.
Receipts
Everything regenerates from one command. Stdlib-only Python: fetchers write immutable, date-stamped raw files; loads are idempotent upserts into SQLite; this page is built from the analysis output. Charts render with Apache ECharts 5.6.0, pinned from cdnjs; their data is embedded in the page.
The reconstruction is where an error would hide, so it got the most engineering. Daily rows and weekly chart rows never mix on a date, which prevents double counting. Dates showing fewer than 20 models get discarded, since archived charts only tracked whatever was popular on snapshot day, a biased sample. From January 2025 the official dataset provides every day; before that, months average 4 to 30 archive sample days, counted in the table below.
The two sources also check each other. On the 95 dates where the archive reconstruction and the official dataset overlap, their estimates of China's share agree within six points on average, despite covering different slices of the model list. And where this series overlaps published snapshots of the recent Chinese share, like Wing's mid-2026 estimate, the numbers land in the same range.
One more piece of provenance. This research was built working with Claude, Anthropic's AI (the Fable 5 model, mostly). It wrote the bulk of the pipeline code and drafted prose under my direction. The questions, the method, the assumptions and the conclusions are mine, and so are any errors.
- Tokens: OpenRouter's official rankings-daily dataset (top 50 models per day, January 2025 onward) plus 119 Wayback Machine snapshots of the rankings page covering November 2023 onward, parsed across every format era.
- Open-weight flag: OpenRouter's catalog marks models with published weights; historical models use author-level defaults with explicit exceptions.
- Downloads: Hugging Face API, top 800 text-generation models by trailing-30-day downloads, derivatives credited via
base_modeltags. - Self-hosted: Ollama's public library, 248 models with cumulative pull counts, classified by model-name prefix.
- Chinese domestic: ModelScope's public API, top 500 text-generation models by downloads, with mirror accounts credited to the original lab.
- Classification: about a hundred lab-to-country mappings, cross-checked against Epoch AI's notable models dataset (CC BY), all public in the repo if you want to audit them. Anonymous "stealth" previews get their own bucket rather than a guess, and the biggest unclassified model is 70 billion tokens in a dataset of hundreds of trillions.
- Shares: each sample day is one observation of traffic composition; months average their days. Ratios within a day are immune to the double counting that token sums would suffer.
Monthly values and sample-day counts
| month | sample days | China | US | Europe+ | community | open share of all |
|---|---|---|---|---|---|---|
| 2023-07 | 3 | 0.0% | 88.0% | 11.9% | 0.0% | 0.2% |
| 2023-08 | 4 | 0.0% | 40.3% | 0.0% | 59.7% | 21.6% |
| 2023-09 | 4 | 0.0% | 12.2% | 0.2% | 87.6% | 67.1% |
| 2023-10 | 5 | 0.0% | 13.9% | 3.6% | 82.5% | 73.1% |
| 2023-11 | 5 | 0.2% | 3.4% | 1.5% | 94.9% | 83.6% |
| 2023-12 | 5 | 2.4% | 9.1% | 7.9% | 80.6% | 80.4% |
| 2024-01 | 5 | 4.0% | 10.4% | 19.9% | 65.6% | 87.5% |
| 2024-02 | 5 | 8.4% | 7.5% | 35.5% | 48.6% | 79.6% |
| 2024-03 | 23 | 7.4% | 6.7% | 54.1% | 31.9% | 67.5% |
| 2024-04 | 30 | 7.0% | 18.7% | 48.3% | 26.1% | 60.9% |
| 2024-05 | 8 | 7.7% | 36.7% | 36.0% | 19.6% | 68.2% |
| 2024-06 | 5 | 7.7% | 42.5% | 25.1% | 24.8% | 63.1% |
| 2024-07 | 6 | 9.9% | 35.1% | 18.6% | 36.4% | 54.3% |
| 2024-08 | 4 | 9.1% | 45.1% | 16.2% | 29.7% | 57.9% |
| 2024-09 | 6 | 7.9% | 53.6% | 15.8% | 22.7% | 45.1% |
| 2024-10 | 5 | 6.7% | 54.3% | 17.8% | 21.1% | 39.8% |
| 2024-11 | 8 | 7.3% | 51.4% | 22.6% | 18.7% | 30.4% |
| 2024-12 | 6 | 14.6% | 45.3% | 22.9% | 17.1% | 27.3% |
| 2025-01 | 31 | 26.7% | 43.1% | 16.1% | 14.1% | 24.5% |
| 2025-02 | 28 | 42.3% | 30.2% | 21.9% | 5.6% | 22.2% |
| 2025-03 | 31 | 51.8% | 33.5% | 12.4% | 2.3% | 24.5% |
| 2025-04 | 30 | 58.5% | 30.9% | 9.3% | 1.3% | 24.2% |
| 2025-05 | 31 | 65.0% | 23.8% | 10.3% | 0.9% | 20.6% |
| 2025-06 | 29 | 72.0% | 17.9% | 10.0% | 0.1% | 24.1% |
| 2025-07 | 30 | 76.0% | 13.8% | 10.2% | 0.0% | 31.9% |
| 2025-08 | 31 | 80.5% | 10.1% | 9.4% | 0.0% | 38.3% |
| 2025-09 | 30 | 72.1% | 20.0% | 7.8% | 0.0% | 23.8% |
| 2025-10 | 31 | 62.4% | 24.9% | 12.7% | 0.0% | 24.7% |
| 2025-11 | 30 | 73.0% | 13.8% | 13.2% | 0.0% | 23.8% |
| 2025-12 | 31 | 64.3% | 19.1% | 16.6% | 0.0% | 33.8% |
| 2026-01 | 31 | 71.0% | 15.0% | 14.0% | 0.0% | 34.9% |
| 2026-02 | 28 | 81.0% | 16.0% | 3.0% | 0.0% | 49.9% |
| 2026-03 | 31 | 84.5% | 14.6% | 0.9% | 0.0% | 54.5% |
| 2026-04 | 30 | 86.1% | 12.7% | 1.2% | 0.0% | 56.2% |
| 2026-05 | 31 | 87.5% | 11.5% | 0.9% | 0.0% | 57.7% |
| 2026-06 | 30 | 90.5% | 8.7% | 0.8% | 0.0% | 62.4% |
| 2026-07 | 31 | 89.3% | 10.0% | 0.6% | 0.0% | 73.4% |
| 2026-08 | 31 | 87.9% | 11.2% | 0.8% | 0.0% | 67.1% |
| 2026-09 | 30 | 88.2% | 10.0% | 1.9% | 0.0% | 70.6% |
| 2026-10 | 13 | 88.2% | 10.5% | 1.3% | 0.0% | 69.6% |
Limitations
- Four windows, not the whole market: direct enterprise API traffic and private clusters stay invisible to every public source. Treat levels as specific to each window; the trend and the agreement are the findings.
- Months are averages over whichever days the archive preserved, 4 to 30 of them.
- Open/closed for dead models relies on author rules; no single minor misclassification moves anything.
- A download is not a deployment. The 30-day window at least keeps legacy models from coasting on history.
- Ollama pulls are cumulative since 2023, which flatters early movers; ModelScope downloads are cumulative too, and its mirror accounts make attribution name-based rather than exact.
Glossary
- Base model
- The original trained model that fine-tunes and quantizations derive from. Credit in this piece always flows back to the base model's lab.
- Cloaked model
- A model offered anonymously (OpenRouter labels these "stealth") so a lab can test demand before announcing. Kept as their own bucket here, since their origin is undisclosed by design.
- Fine-tune
- A model made by further training an existing one on new data. The 2023 community era was almost entirely fine-tunes of Meta's Llama models.
- Inference
- Running a trained model to produce output, as opposed to training it. Token counts measure inference.
- OpenRouter
- An aggregator that routes developers' requests to hundreds of models across many providers through one API, and publishes model-level usage rankings. The only public, model-level, multi-provider usage data there is.
- Open weights
- A model whose trained parameters anyone can download and run. Not the same as open source: training data and code usually stay private, and licenses vary. The opposite is a closed model, reachable only through its maker's API.
- Self-hosted
- Running a model on hardware you control, from a laptop to a private cluster, rather than calling someone's API. Ollama is the most popular tool for doing this locally, which makes its pull counts a window into that population.
- Quantization
- Compressing a model's weights to lower precision so it runs on smaller hardware. Quantized copies are usually republished by third-party accounts, which is why raw download counts misattribute so much volume.
- Token
- The unit models read and write text in, roughly three quarters of a word. Token counts are the standard measure of model usage.