China's open-model wave has a new flagship. Tencent released Hy3 — the third generation of its Hunyuan model — on July 6, 2026, a 295-billion-parameter system it has fully open-sourced under the permissive Apache 2.0 license.
The architecture
Hy3 is a mixture-of-experts model with 295B total parameters but only 21B active per token, giving it near-frontier quality at a fraction of the inference cost. It ships a 256K-token context window and three selectable inference modes that blend fast responses with slower, deliberate reasoning.
The benchmarks
On Tencent's numbers, Hy3 scores 90.4 on GPQA Diamond, approaching GPT-5.5's 93.6, and 84.2 on BrowseComp, matching GPT-5.5 on agentic search. Its HLE-with-tools score of 53.2 tops DeepSeek-V4-Pro's 48.2, and it posts 78.0 on SWE-bench Verified and 57.9 on SWE-bench Pro. Internal testing cut its hallucination rate from 12.5% to 5.4%. VentureBeat's read: it beats Zhipu's GLM-5.2 almost everywhere except coding — at half the parameter count.
The license and the giveaway
The aggressive part is the licensing. Apache 2.0 imposes no commercial restrictions, unlike the bespoke licenses some rivals attach. Tencent is also giving the model away for now: free through July 21 via OpenRouter's tencent/hy3:free endpoint, after which standard pricing applies — a continuation of its high-cost-efficiency strategy.
The bigger picture
Hy3 compounds a shift already visible on OpenRouter, where Chinese-origin models now move a large share of developer traffic. It joins Qwen, DeepSeek, GLM and Xiaomi's MiMo in the open-weight push — though, unlike the closed-weight MiMo, Hy3 ships its weights. The usual caveats apply: the benchmarks are self-reported, and Chinese models typically carry political-topic restrictions that enterprise buyers weigh.