Chinese AI Models Cross 60% Market Share on OpenRouter

Chinese-built large language models now account for more than 60% of all token traffic on OpenRouter, the largest third-party AI model router used by developers. The shift, tracked across recent platform data, marks a roughly 50x increase in Chinese model share over the past 18 months and is the clearest single signal that the open-weights frontier has effectively become Chinese-led.

The dominant names — Kimi K2.6 (Moonshot AI), DeepSeek V4, GLM-5.1 (Zhipu AI), and Qwen 3 (Alibaba) — are now the default choice for an entire generation of developers building agentic, coding, and high-volume inference workloads. Meta's delayed Avocado model, the last credible US open-weights frontier candidate, has gone silent.

What the Data Shows

OpenRouter is the most-watched proxy for real-world third-party LLM usage. It routes API traffic across more than 300 models and publishes token volume rankings that capture actual developer adoption rather than benchmark scores or marketing claims.

The shift has been dramatic. In October 2024, Chinese-developed models accounted for approximately 1.2% of all OpenRouter token volume. By March 2025, that figure crossed 10%. By Q3 2025, it passed 25%. By April 2026, Chinese models collectively processed over 45% of all tokens on the platform. The latest reads put the figure above 60%.

The leading models by recent token volume:

MiniMax M2.5 topped weekly leaderboards with 2.45 trillion tokens consumed in a single week — a 197% week-over-week jump. Moonshot AI's Kimi K2.5 followed at 1.21 trillion tokens. Zhipu's GLM-5 ranked third at 780 billion tokens, up 158% week-over-week. DeepSeek V3.2 held roughly 5.6% platform-wide share. MiMo-V2-Pro and Qwen 3.6 Plus together accounted for roughly 49% of all coding-specific token volume.

Why This Happened

Three forces compounded.

First, capability. Chinese open-weight models are no longer chasing the closed-model frontier — they're competing with it head-to-head on benchmarks that matter. Kimi K2.6 hit 58.6% on SWE-bench Pro (the agentic coding benchmark), within 6 points of Claude Opus 4.7. GLM-5 from Zhipu AI scored 85 on the BenchLM composite and 77.8% on SWE-bench Verified. DeepSeek V4 introduced major reasoning and agentic upgrades on April 24 with full weights released on Hugging Face.

Second, price. Chinese models are 10–20x cheaper than comparable US frontier models depending on workload. Kimi K2.6 lists at $0.60 per million input tokens and $2.50 per million output — roughly 8x cheaper than Claude Opus 4.7 for comparable coding-agent tasks. For developers running high-volume agentic workloads where each task consumes hundreds of thousands of tokens, the cost gap is the entire business case.

Third, openness. All four leaders ship full weights, not just API access. That matters for enterprise customers who can't move regulated data to a US-hosted API, for hobbyists running models locally, and for downstream builders fine-tuning specialized variants. The closed-model labs have systematically refused to open-weight their frontier; the Chinese labs have systematically released theirs.

Why This Matters

The open-weights tier was a US-led category until late 2024. Llama 2, Llama 3, Mistral 7B, Falcon, and the early Mixtral variants all came out of Western labs. Today, Meta's open-weight roadmap has stalled (Llama 4 underperformed, the next Avocado model is unreleased), Mistral has shifted toward closed enterprise products, and the only credible Western open-weight frontier release in 2026 has been Gemma 3 from Google — which sits below Kimi K2.6 and DeepSeek V4 on most benchmarks.

The strategic consequence: every developer building on open-weight models in 2026 is, by default, building on Chinese models. That has practical implications. Data routed through these models may pass through Chinese-hosted endpoints (when using API providers based in China). Fine-tuning workflows and downstream applications inherit any biases, safety calibrations, or content restrictions baked into the base model. The supply chain for the entire open-weight stack now depends on Beijing-based labs continuing to release frontier weights — a dependency that didn't exist 18 months ago.

For US labs, the competitive question has changed shape. Anthropic, OpenAI, and Google can compete with closed-model APIs for enterprise contracts. They cannot compete in the open-weight tier, because they don't ship weights. That's now a structural advantage Chinese labs hold and can extend.

What Industry Watchers Are Saying

The shift has been building for months but the 60% threshold is producing fresh commentary.

Dataconomy noted in February that the 61% level "reflects what developers actually choose when cost and capability are weighted together — not what marketing teams pitch." That framing has held up: the OpenRouter data is a revealed-preference metric, not a stated-preference survey.

US AI policy analysts have responded with concern. The US-China Economic and Security Review Commission's recent report on "Two Loops: How China's Open AI Strategy Reinforces Its Industrial Dominance" framed the shift as a deliberate strategic play by Chinese tech to capture the developer tier as a long-term industrial leverage point.

The counter-read: this is fundamentally a price story, not a politics story. Western developers aren't choosing Chinese models because they're Chinese — they're choosing them because they're 10x cheaper at parity. If US labs released competitive open-weight models at competitive prices, the trend would reverse. None has.

What's Next

Three things to watch.

First, whether Meta finally ships the next Avocado open-weight model. If the release lands and it's competitive on benchmarks, the Chinese share will compress meaningfully. If Meta keeps slipping, the trend extends.

Second, OpenAI's open-weight strategy. Sam Altman has publicly committed to releasing an open-weight reasoning model "in the coming months" but the release has been pushed multiple times. Whether OpenAI actually ships one — and at what scale — will determine whether US labs retake any share at all.

Third, regulatory response. US government concern about Chinese open-weight model adoption is real. Some form of export-control-style restriction on US developers using Chinese-hosted inference endpoints has been floated. If it lands, the market dynamics change overnight.

Bottom Line

The 60% threshold is symbolic but the underlying trend is real. The open-weights frontier is Chinese-led, the cost gap is structural, and US labs have either chosen not to compete in this tier or are too slow to do so. For developers building cost-sensitive agentic workloads in 2026, the default stack is now Chinese weights running on a Western GPU. That's a quietly enormous shift in the AI supply chain — and the labs that haven't priced it in yet are about to.