Hy3

Hy3

Quick answer: Hy3 is Tencent's third-generation flagship Hunyuan language model, released 7 July 2026. Built end-to-end by Tencent, it is positioned as "business-savvy and well-rounded" with outstanding capabilities in coding, long-form content generation, and agentic task execution. On LHTB it scored 0.288 mean reward (rank 14) at approximately $2.50 per task — the lowest cost of any model evaluated.

At a Glance

Where Hy3 leads

• Lowest cost per LHTB task (~$2.50) among all 18 evaluated models.
• Strong agentic capabilities designed for real enterprise workflows.
• Deep integration across Tencent's ecosystem (Tencent Docs, QQ, WeChat, Tencent Video).

Where it lags

• LHTB mean reward of 0.288 places it in the lower third of the frontier field.
• Primarily targeted at Tencent's domestic ecosystem; international developer access may vary.

Best for: Cost-efficient agentic pipelines, business automation workflows, and Tencent ecosystem integrations.

What Hy3 Is

Hy3 (Hunyuan 3) is the latest iteration of Tencent's internal LLM program, built entirely end-to-end by the Tencent Hy team. The model is the third full-generation successor in the Hunyuan series and succeeds the Hy2 family. Tencent describes it as excelling at three areas: coding (fast and stable engineering execution), long-form content (coherent, sustained text production), and agentic capability (autonomous task completion aligned with real workflows).

The model is designed for deep integration into Tencent's product ecosystem. It powers multiple Tencent products including Tencent Docs, QQ Browser, QQ, Sogou Input Method, Tencent Video, and WeChat Official Accounts — reflecting a strong production co-design focus. Tencent AI Studio (Hy AI Studio) provides developer access alongside a broader Yuanbao consumer product.

On LHTB — which benchmarks 18 frontier models on 46 hard, 90-minute terminal tasks — Hy3 scored 0.288 mean reward and solved 1 of 46 tasks outright, placing 14th overall. Notably, LHTB reports a per-task cost of approximately $2.50 for Hy3, the lowest cost of any model in the evaluation, making it the most cost-accessible option in the benchmark field.

Specifications

Related Benchmarks

• Long-Horizon Terminal-Bench (LHTB) — 0.288 mean reward, rank 14 (1/46 solved) · ~$2.50/task