Tencent Hunyuan Hy3: Specs, Pricing, Benchmarks, and Everything You Need to Know

License

The model is released under the Tencent Hunyuan Community License — a non-exclusive, royalty-free, non-transferable limited license allowing use, reproduction, distribution, and derivative works for the "Territory" (details in the license file) .

Pricing

• Pay-as-you-go API pricing: $0.063 per million input tokens and $0.210 per million output tokens; cached tokens at $0.021 per million .
• On launch, OpenRouter offered Hy3 Preview as a fully free endpoint (no credit card, no token metering) for about two weeks; that free period has since ended .
• Tencent considers this among the cheapest capable LLMs globally .

Availability Platforms

• Open-source weights on GitHub (Tencent-Hunyuan/Hy3-preview) and Hugging Face .
• API endpoints via Tencent Cloud's TokenHub platform and OpenRouter .
• Serving frameworks: vLLM, with official deployment recipes .
• Configurable reasoning levels (disabled, low, high) to trade off latency vs. depth .

Performance Improvements

• Hallucination rate: Down by more than half compared to Hy3 Preview, based on real-log evaluations . In a specific game-assistant integration (Path of Exile), hallucination dropped from 4.5% to 2.8% .
• General knowledge error rate: Halved relative to Preview .
• Agent task benchmarks:
  • ClawEval pass³: 68.5 — beats DeepSeek V4 Pro (62.4) and Qwen 3.7 Max (65.2) .
  • SkillsBench: 55.3 (up from 29.1 in Preview) — also leads DeepSeek V4 Pro and Qwen 3.7 Max .
  • BrowseComp: 84.2 — nearly ties GPT 5.5 (84.4) .
  • MathArena Apex: 38.7 (up from 12.8 in Preview) .
  • SWE-bench Verified: 74.4% .
  • Terminal-Bench 2.0: 54.4% .
• Agent & coding core capabilities improved 20–30% overall . Co-design testing indicates ≥90% of Agent tasks can be successfully executed with Hy3 across data processing, document generation, research reports, web page creation, and life-decision scenarios .
• Internal blind test (270 Tencent engineers on real work tasks): Hy3 scored 2.67/4 vs. GLM 5.1's 2.51/4 .
• Multi-agent orchestration: Marvis (Tencent's IT agent) reached 93.7% task completion, 92% correct dispatch across 6 collaborating agents .

Development Timeline

• February 2026: Tencent overhauled its pre-training and reinforcement-learning frameworks from the ground up .
• ~6 weeks later: Hy3 Preview training began .
• April 21–23, 2026: Hy3 Preview released and open-sourced on GitHub .
• July 6, 2026: Production Hy3 officially launched with full ecosystem integration .

Ecosystem Integrations

• Agent frameworks: Integrates with OpenClaw, OpenCode, and KiloCode .
• Tencent products:
  • Yuanbao (Tencent's AI assistant) — gained file-delivery capability; common-sense error and hallucination rates both fell more than half .
  • ima — Agent system stability reached 95.1%; knowledge-base QA reasoning quality improved nearly 19% .
  • Marvis (internal IT agent) — core scenario task completion 93.7% .
  • CodeBuddy & WorkBuddy — 54% lower TTFT, 47% lower end-to-end latency, >99.99% success rate .
  • WeGame / Path of Exile AI assistant — multi-turn reasoning + tool dispatch success rate 92%, hallucination down from 4.5% to 2.8% .
• WeChat & gaming — Blog AI and game assistants also benefit .

Competitive Positioning Among Chinese Open-Source AI Labs

Hy3 positions itself as a practical, agent-first model rather than a pure parameter-size play. Key competitive points:

• Agent differentiation: Tencent explicitly targets agentic and coding workloads where it claims to match or beat larger rivals (e.g., Kimi-K2.5 at 1T+ parameters) at a fraction of the cost . On ClawEval pass³ and SkillsBench, Hy3 outperforms DeepSeek V4 Pro and Qwen 3.7 Max .
• Pricing leadership: DeepSeek and Qwen are also known for aggressive pricing, but Hy3's sub-$0.07/1M input is among the most competitive .
• STEM reasoning: Strong on FrontierScience-Olympiad, IMOAnswerBench, Tsinghua math-PhD qualifying exam, and China High School Biology Olympiad — comparable to DeepSeek and Qwen's frontier models .
• Weakness: Some third-party aggregate benchmarks (e.g., Artificial Analysis GDPval-AA) show Hy3 Preview trailing open-weight peers on certain real-world agentic work tasks , though production Hy3 appears to have closed much of that gap.
• Overall lab positioning: Tencent joins DeepSeek, Alibaba (Qwen), Zhipu AI (GLM), and Baidu (ERNIE) as a major Chinese open-source AI lab. Hy3 is Tencent's most credible entry yet into the "frontier open-weight" race, betting on agent orchestration and low cost rather than raw parameter scale.