A model most Western users had never heard of quietly became the most-called large model (大模型) on one of the world's biggest open-model platforms. Within a week of its July 2026 release, Tencent's Hunyuan Hy3 (混元 Hy3) was generating 68 times more API traffic than its predecessor — and sitting at number one on OpenRouter's global usage leaderboard.
That is not a benchmark win. It is a "people actually deployed it" win.
From preview to global rollout
Hy3 first appeared as an open-source preview in late April 2026, after Tencent rebuilt its pretraining and reinforcement-learning infrastructure earlier that year. The full model was released on July 6, 2026, and by August Tencent had opened it to international users through its WorkBuddy agent workspace, its Design Miora (妙境) creative studio, and Tencent Cloud's TokenHub service.
The speed of that rollout matters. Tencent did not treat Hy3 as a research curiosity; it treated the model as infrastructure to be called, embedded, and billed by the token.
What Hy3 actually is
Hy3 is the first model from Tencent's rebuilt training stack. The headline numbers:
- A mixture-of-experts (混合专家, MoE) design with 295 billion total parameters, but only 21 billion activated per token — less than 7% of the weights fire on any given request.
- A 256K-token context window, roughly 200,000 words of working memory.
- A "fast-slow thinking" (快慢思考融合) architecture that blends quick replies with deeper reasoning passes.
- Native agent (智能体) tool use, so it can run code, browse, and manipulate files inside a workflow rather than just chat.
The sparse design is the whole point. Because only 21B parameters activate, the cost per call collapses toward that of a much smaller dense model — which is exactly why a price-sensitive platform like OpenRouter rewarded it with traffic.
Why developers reached for it
Tencent did not just drop weights and walk away. Hy3 shipped with a real distribution story:
- An open-weight (开源权重) release under the Apache 2.0 licence on Hugging Face and ModelScope.
- Same-day integration into third-party agent frameworks: Hermes, Kilo, Cline, OpenClaw, OpenCode, and Cherry Studio.
- Access through Tencent's own products: the WorkBuddy agent workspace, the Design Miora (妙境) creative studio, and the TokenHub model-as-a-service platform on Tencent Cloud.
Tencent reported that inside its own workplace tests, Hy3 cleared 90%+ task success in WorkBuddy and cut average task-completion time by 34% versus the previous Hy model. Those are the company's own numbers, but they point at the strategy: win on task completion and price, not on a leaderboard line.
The price math
On Tencent Cloud TokenHub, Hy3 is priced at roughly ¥1.2 per million input tokens and ¥4 per million output tokens (cached input as low as ¥0.4), with a personal plan around ¥28 per month — about US$0.17 / HK$1.3 per million input tokens and US$0.56 / HK$4.4 per million output tokens, or roughly US$3.9 / HK$30.6 a month for the personal tier. On OpenRouter, listings started near US$0.13 input / US$0.53 output per million tokens.
Either way, the message is the same: this is a model built to be called a lot, cheaply.
Beyond the headline
Hy3's OpenRouter (全球开源模型平台) climb is worth reading as a signpost rather than a coronation. For years, the open-model conversation outside China leaned on US and European names — Meta's Llama, Mistral's family, and a rotating cast of research releases. A Chinese model taking the top usage slot signals that developers now choose on cost and tooling fit, not on a lab's home address. That is healthy for buyers: it forces every vendor to compete on price-per-task.
For a reader outside China, the practical takeaway is access. Because the weights are open and the integrations are mainstream, you can run the same model a Shenzhen team runs without touching a China-only service. The flip side is ranking concentration: when one platform's leaderboard becomes a proxy for quality, niche but better-suited models can be overlooked. Treat the chart as one signal among many, not a verdict.
Trend reading
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the real test for Chinese large models (大模型) in 2026 is whether open-weight systems can hold up inside real enterprise workflows, not just top the open leaderboards.
Honest limitations
- The 68× traffic jump and #1 ranking describe a short launch window; a sustained position depends on whether developers keep calling Hy3 after the novelty fades.
- The 90%+ success rate and 34% time savings are Tencent's internal claims, not independently audited benchmarks.
- Hy3 is strong on agent and coding tasks but, by Tencent's own framing, is positioned as the pragmatic option rather than the absolute frontier model.
- Documented Western enterprise adoption is still early; most visible usage sits inside Tencent's ecosystem and adjacent open-source tooling.
What readers can do now
- Try it free where it is open. Pull the Apache 2.0 weights from Hugging Face or ModelScope, or test Hy3 inside WorkBuddy (free through Aug 31, 2026 PT) before paying for tokens.
- Benchmark it on your own task, not a leaderboard. Hand it a real multi-step workflow — a spreadsheet cleanup or a code refactor — and compare cost-per-completed-task against a closed API.
- Watch the agent ecosystem, not just the model. Hy3's edge comes from OpenClaw, OpenCode, and Cline support; evaluate those frameworks alongside the model if you plan production use.