Tencent released Hy3 on July 6, a 295-billion-parameter Mixture-of-Experts model that activates 21 billion parameters per token, plus 3.8 billion parameters in a multi-token-prediction (MTP) layer, with a 256K-token context window. The MoE design means only the 21-billion active path runs per token instead of the full 295 billion, which is what keeps inference cost and latency close to a mid-sized dense model despite the total parameter count. The MTP layer lets Hy3 predict more than one future token per forward pass, a technique labs use to speed up generation without retraining a separate draft model. Tencent says the release outperforms similarly sized models and closes ground on open-source flagships built with two to five times as many parameters, crediting a stronger post-training data set than its earlier Preview version used.

Simon Willison, whose blog tracks model releases closely, surfaced Hy3 the same day. He ran his usual smoke test: asking the model to draw an SVG of a pelican riding a bicycle. Hy3 produced a flat-style cartoon rendering. Willison offered light commentary and no formal benchmark run, relying on Tencent’s own reported numbers rather than an independent evaluation.

That gap matters for anyone reading the release as validated. The two-to-five-times comparison is Tencent’s own framing, not a third-party measurement. Willison’s pelican-and-bicycle SVG check tests whether a model can follow one illustration prompt, not whether it matches larger open flagships. No outside lab or public leaderboard has confirmed Tencent’s parity claim, so treat it as a vendor statement until independent benchmarks land.

The weights are heavy. Full precision files total 598GB on Hugging Face, and an FP8 quantized build cuts that to roughly 300GB, still large enough to need multi-GPU serving for most teams. That footprint, combined with the 21-billion-parameter active path, points Hy3 at organizations already running MoE inference infrastructure rather than at solo developers testing on a single card.

Tencent is making Hy3 available on OpenRouter at no cost through July 21. Two weeks of free routing access functions as a distribution push, not a pricing decision: it gets developers running side-by-side comparisons against Llama, Qwen, and DeepSeek before Tencent likely reintroduces metered pricing, converting free trial usage into adoption data before anyone commits budget to the model.

Hy3 extends a run of large open-weight releases out of Chinese labs, following DeepSeek’s V3 and Alibaba’s Qwen line. Tencent joining that trend widens the field of strong open-weight bases for teams picking one, though each entry still carries vendor benchmark claims that outside researchers have historically taken weeks to verify or dispute.

Teams evaluating Hy3 for production should treat the free OpenRouter window as a benchmarking opportunity rather than a launch signal. Run your own evals against the tasks you actually care about before July 21 closes the free access, since Tencent’s comparison numbers have no independent confirmation yet.

Simon Willison reported Tencent’s Hy3 release on his blog on July 6, 2026, citing the company’s own stated benchmarks.