In April 2026, Tencent Hunyuan released Hy3 Preview — the debut of Shunyu Yao (lead author of the ReAct paper) since joining Tencent. Back then the model could handle some real requests, but in hands-on coding tests it still trailed the flagships of the same period. Two months later, the official Hy3 release is here: substantive gains in agent and coding ability, benchmark scores now in the neighborhood of GLM-5.2, and office-scenario task success up from 72% to 90%. If you're looking for a cost-effective homegrown coding agent model, Hy3 is now worth a try — and it's still free for a limited time inside WorkBuddy.

What the official Hy3 release is

Tencent Hunyuan Hy3 (Hunyuan 3) official version is the flagship large language model released in July 2026, with agent and coding capabilities as its headline.

Core specs (from Tencent Cloud's site, cloud.tencent.com/product/tclm):

• Architecture: MoE (mixture of experts), 295B total parameters, 21B activated
• Context: 192K max input, 128K max output
• Thinking modes: three tiers — no_think, think_low, think_high
• Positioning: an official release polished for real business scenarios

Hands-on: a leap in coding agent ability

According to APPSO's hands-on tests, the official Hy3 shows clearly visible improvement over Preview in frontend web generation. The test cases included complex 3D web simulations, strict game logic, and multi-framework/library calls.

Golden Gate Bridge 3D experience: with the same prompt, Hy3 designed several camera modes — close-range fly-through, mid-range orbit, high aerial, auto-orbit, and cinematic tour — and the bridge deck, cars, and water reflections all came across as more realistic than Preview. The Preview version delivered "an unfinished work-in-progress, rough and limited."

Game remakes: multiple game remake tasks previously used to test GLM-5.2 (Minecraft, Black Myth: Wukong, Subway Surfers) now mostly reach the same level with Hy3, with some doing even better. In the Black Myth: Wukong remake, GLM-5.2 and Opus 4.8 leaned 2D or dark-style, while Hy3 delivered a fairly complete 3D game.

Code project analysis: hand Hy3 the code of a classic paper from GitHub and ask it to find bugs and optimization directions, and it proactively delivered a detailed Markdown report clearly listing the project's research content, confirmed bugs, potential robustness risks, and performance and architecture optimization directions. It could also convert a Matlab project to another language, convert a Python project to a Google Colab ipynb, and fix the bugs it discovered itself.

Benchmark performance

Hy3 is arguably the model that has improved the most since its preview on benchmarks like SWE-bench, HLE, and BrowseComp. In official comparisons, Hy3's main reference points are DeepSeek V4 Pro, GLM 5.2, Qwen 3.7 Max, GPT 5.5, Opus 4.8, and Doubao Seed 2.1 Pro; overall it lands near GLM 5.2, exceeding some of the others on certain dimensions.

How to use it

Hy3 is available on two platforms:

• Tencent Yuanbao: pick fast-thinking mode, enter a clear request stating the file type you need (PPT, Word, Excel, PDF, HTML, etc.), and Yuanbao will automatically trigger its agent capabilities
• WorkBuddy (Tencent's all-scenario AI workspace): similar to Codex, it can start directly from local files to handle computer tasks; Hy3 is free there for a limited time

In WorkBuddy testing, Hy3 showed strengthened task planning, long-document information processing, and multi-tool calling when facing complex workflows, with office-scenario task success rising from 72% to 90%.

API pricing

Hy3's API pricing is highly competitive among homegrown flagships:

Input pricing is one-third of DeepSeek V4 Pro's. On cache-hit pricing DeepSeek is lower, but on standard input/output Hy3 is cheaper.

💡 Tip: Tencent Cloud together with Hunyuan also launched a Hy3 preview Token Plan package, starting at ¥28/month for the personal edition — a fit for agent development and long-term use.

Who it's for

• Frontend developers: Hy3's delivery quality on 3D web pages, game remakes, and SVG animation approaches GLM-5.2 and GPT-5.5
• Data / research workers: reasoning and complex STEM reasoning have improved significantly, enough for advanced data analysis and scientific research
• Agent developers: low API pricing, long context (192K input), and high task success — a fit for building production-grade agents
• Office users: use it directly through Yuanbao and WorkBuddy with no coding required

Still worth watching

Tencent itself admits the cadence is "an update every two months," and some models iterate even faster than that. Once GPT-5.6 opens up fully and Gemini 3.5 Pro launches, the benchmark becomes a new bar all over again. Hy3 completed a clearly visible iteration this time — but whether it can keep up this update pace matters more than this generation of the model itself.