Tencent Ships Apache-Licensed Hy3, Beats GLM-5.2 Except on Coding

Tencent quietly moved the goalposts on Chinese open-weights this week, promoting Hunyuan Hy3 out of preview and shipping it under Apache 2.0, as VentureBeat reported. The framing that stood out is not the parameter count, it is the qualifier in the headline: Hy3 reportedly beats Zhipu's GLM-5.2 on most tasks and loses on coding. When a Chinese lab willingly concedes the coding leaderboard, that is a real signal about where the competition is now differentiated.

The technical shape is a 295B-parameter mixture-of-experts with 21B active per token, 256K context, and top-8 routing across 192 experts plus a shared always-on expert per layer. That is a well-studied MoE recipe rather than an exotic one, and the point is efficiency: Tencent is pricing the API at 1 yuan per million input tokens, 4 yuan per million output tokens, and just 0.25 yuan per million when cache hits. At that level the cost story starts to matter more than the benchmark story for anyone deploying agents at volume.

The most concrete claim is internal but worth taking at face value with a caveat. On WorkBuddy, Tencent's enterprise assistant, task resolution reportedly climbed from 72% on the April preview to 90% on the official release, with average time to complete a task falling by 34%, according to Pandaily's writeup. Gains on SkillsBench (29.1 to 55.3) and MathArena Apex (12.8 to 38.7) point in the same direction, and hallucination rates are reported to have roughly halved. Hy3 is already being wired into CodeBuddy, Yuanbao, ima, QQ Browser, and Tencent News, with more products queued behind them.

The honest caveat is that Tencent's published benchmark comparisons appear to have leaned on DeepSeek-V3 and GLM-4.5 as baselines rather than the newer V4 or GLM-5.2, so the tidy "beats GLM-5.2 everywhere except coding" line is doing a lot of work that the underlying charts do not fully back up. The reporting also does not quantify the coding gap, so treat that as a known unknown rather than a resolved question.

What is genuinely useful for a leader watching this space: an Apache-licensed model at these prices makes it harder for anyone selling a Chinese frontier API to hold margin, and it gives global teams building agent stacks another self-hostable option that is not Qwen or DeepSeek. If the WorkBuddy gains hold up outside Tencent's own product, the interesting fight in 2026 is less about who leads the leaderboard and more about who can afford to run at leaderboard quality.