The Apache 2.0 Mixture of Experts model activates 21B parameters per pass, weighs 598 GB in full precision and 300 GB in FP8, and runs free on OpenRouter through July 21, 2026.
Tencent's Hy Team released Hy3 on July 6, 2026, an open-weights Mixture-of-Experts language model with 295 billion total parameters, 21 billion active per pass, and a 3.8 billion-parameter Multi-Token Prediction layer dedicated to outputting multiple tokens at once. The weights are on Hugging Face under Apache 2.0, the full model is 598 GB and an FP8 quantisation is 300 GB, context length is 256K, and OpenRouter is hosting it free through July 21, 2026.
Hy3 adds a 3.8-billion-parameter Multi-Token Prediction head to its decoder. Standard transformers emit one token per forward pass and feed it back in for the next. The MTP head predicts several candidate tokens at once, and the decoder commits one of them while the rest can be verified or discarded in a single pass. The 21-billion-parameter active pathway therefore amortises its compute across more output tokens per pass than a vanilla decoder would. Tencent documents the design in its release post and on the Tencent-Hunyuan/Hy3 GitHub repository.
The active-parameter count, not the total, drives the GPU memory bandwidth and FLOPs bill per generated token. A 295B MoE with 21B active is closer in inference cost to other ~20B-class active models than to dense 295B peers, and the MTP layer is the mechanism Tencent is using to push that figure lower. The trade-off is that the 3.8B MTP parameters still need to be loaded into GPU memory alongside the active experts at every forward pass, which is why the FP8 quantisation at 300 GB is the practical choice for self-hosted serving at 256K context. Throughput on a fixed GPU budget becomes a function of how often the MTP head's full prediction is committed rather than rejected, and that is the deployment-level question no benchmark can fully resolve.
Tencent says Hy3 outperforms similar-size open models and rivals flagship open-source models with 2-5x the parameter count. That is a vendor claim, repeated through Simon Willison's link blog, and it has not been independently benchmarked at release. Willison's own quick prompt, asking Hy3 to "Generate an SVG of a pelican riding a bicycle", is a vibe check rather than a capability evaluation, and reads more as colour than as evidence. Independent numbers on MMLU, MMLU-Pro, GPQA, SWE-bench, and the standard agent-evaluation suites will determine whether the 2-5x framing survives contact with third-party runs.
Tencent positioned Hy3 around agent capabilities and product integration. Post-training was scaled up after the Hy3 Preview in late April 2026, using feedback collected from more than 50 internal Tencent products. The plan is for the model to be useful inside Tencent's own applications first, with the public release following the internal tuning cycle. The Tencent Hy portal is the company's central distribution point.
For developers who want to evaluate Hy3 without standing up their own serving stack, the practical hook is the free OpenRouter window through July 21, 2026. After that, hosted access returns to standard OpenRouter pricing, and self-hosting requires either the full Hy3 weights at 598 GB or the FP8 quantisation at 300 GB, plus serving infrastructure sized for a 256K-context MoE. The Apache 2.0 license means commercial use and derivative fine-tunes are unrestricted.
The free OpenRouter window closes July 21, 2026. Independent MMLU, GPQA, and SWE-bench scores from third-party labs are expected before then, and those numbers, rather than Tencent's release language, will determine whether the 2-5x claim survives contact with the rest of the open-weights field.