Tencent Hy3 Milestones: A 295B MoE That Grew 68x and Topped OpenRouter in Its First Week

Tencent's Hy3 AI model has achieved a remarkable set of milestones in its first two weeks since launching on July 6, 2026. Built on a Mixture-of-Experts (MoE) architecture that activates only 21 billion of its 295 billion total parameters per token, Hy3 has posted 68x call volume growth over its predecessor Hy2, claimed the #1 spot on OpenRouter's global model call volume ranking, and generated so much demand that it temporarily overloaded Tencent's own compute infrastructure . Here is the detailed, source-cited breakdown of what Hy3 has accomplished and how it got there.

Architecture & Specifications

Hy3 is a Mixture-of-Experts (MoE) model with 192 experts and top-8 routing . Key specifications from the official Hugging Face model card and Tencent's announcement :

Specification	Value
Total parameters	295 billion
Active parameters per token	21 billion
MTP layer parameters	3.8 billion
Number of layers	80 transformer layers + 1 Multi-Token Prediction layer
Attention heads	64
Context window	256K tokens (262,144)
Architecture	Dense-attention, sparse-FFN MoE

The model supports configurable reasoning modes — a direct "no-think" mode plus low/high chain-of-thought modes for complex tasks . Tencent describes it as a "hybrid fast-and-slow-thinking model" .

API Call Volume Growth vs. Hy2

Within its first week, Hy3's total API call volume grew more than 68× compared to Hy2, the previous-generation model . A Chinese-language report from Sina Finance notes that Hy3's growth trajectory "continued the upward trend of the Hy3 preview version but with a steeper slope" .

OpenRouter Leaderboard Performance

• #1 in global model call volume on OpenRouter within one week of launch .
• Three days after launch, Hy3 rose to 8th place on OpenRouter's most-popular model ranking during its limited free-access phase .
• The Hy3 Preview had already held the #1 spot on OpenRouter's usage rankings for three consecutive weeks before the full release .

Token Processing Volume & Infrastructure Strain

Exact total token volume figures beyond the weekly OpenRouter tallies were not published in the sources reviewed, but the demand surge was dramatic enough to overload Tencent's compute infrastructure. On July 8 (two days after launch), WorkBuddy's inference computing resource consumption peaked, with afternoon queue rates exceeding 50% . Tencent had to urgently allocate additional compute capacity, announcing service restoration on the morning of July 9 .

The Hy3 Preview had already accumulated 7.7 trillion tokens on OpenRouter between April 23 and May 12, according to Tencent's Q1 2026 earnings materials .

WorkBuddy Ecosystem Adoption

• 60% of WorkBuddy users who chose a custom model selected Hy3 .
• WorkBuddy itself has grown to 13 million DAU and 20 million MAU .
• After Hy3 integration, WorkBuddy's internal task success rate increased from 72% to 90%, and average execution time fell by 34% .
• Tencent's Yuanbao (AI assistant) saw hallucination rates in long-document and AI search scenarios drop by more than half .

Pricing

• $0.14 per million input tokens via OpenRouter
• $0.58 per million output tokens via OpenRouter
• OpenRouter reports that prompt caching can reduce effective cost by 60–80%
• China Merchants Securities reported the API input price at RMB 1 per unit on Tencent Cloud's TokenHub platform

Open-Source Licensing

Hy3 is released under the Apache 2.0 license with no geographic restrictions . The full model weights are available on Hugging Face at tencent/Hy3 .

Integration Across Tencent Products

Hy3 has been deployed across multiple Tencent products :

• WorkBuddy (enterprise intelligent workplace assistant)
• CodeBuddy (developer coding assistant)
• Yuanbao (consumer AI assistant)
• ima, Marvis, QQ Browser, Tencent News, WeGame, Sogou Input and nearly 50 more products in the pipeline

Positioning Strategy: Agents Over Model Size

Tencent's strategy with Hy3 is explicitly about agents over model size. As Forbes covering the launch noted, Tencent is betting that an efficiently activated MoE model (21B active of 295B total) at dramatically lower cost can rival much larger dense-model flagships . Key benchmarks cited include:

Benchmark	Hy3 Score
SWE-bench Verified	74.4% (Preview)
BrowseComp	84.2 (best open-weight search agent score)
DeepSearchQA	91.0
MCP-Atlas	79.1 (best open-weight tool orchestration)
AA-LCR (long-context retrieval)	73.4

Trajectory from Hy3 Preview

The Hy3 Preview launched April 23, 2026 as the first model built on Tencent's completely rebuilt pre-training infrastructure . It was described as the "first model trained on Tencent's rebuilt infrastructure" and delivered significant improvements over Hy2 in complex reasoning, instruction-following, in-context learning, code generation, and agent capabilities . The Preview held OpenRouter's top usage slot for three consecutive weeks , and the full release amplified that momentum — with total calls growing 68× over Hy2 versus an already strong Preview growth rate . Between Preview and GA, Tencent's team refined the model through global developer feedback and across its own massive product ecosystem, with measurable improvements: hallucinations dropped from 12.5% to 5.4% and commonsense reasoning errors fell from 25.4% to 12.7% .

Key takeaway: In its first ~10 days, Hy3 achieved 68× call volume growth vs. Hy2, became the #1 model by call volume on OpenRouter, overloaded Tencent's own compute infrastructure due to demand, captured 60% of WorkBuddy custom-model users, and demonstrated enterprise agentic improvements (90% task success rate, 34% faster execution). This builds directly on the Hy3 Preview's momentum, which had already dominated OpenRouter rankings since late April.