Tencent Releases Open 295B Mixture-of-Experts Model Hy3

Tencent's Hy team has released Hy3, a 295B-parameter Mixture-of-Experts (MoE) model that activates only 21B parameters per token. The model is aimed at reasoning, agentic workflows, and long-context tasks, and its weights are available under the Apache License 2.0. Hy3's architecture features a sparse MoE with 192 experts and top-8 routing, where only 8 experts fire per token, keeping compute low.

The model also uses a Multi-Token Prediction (MTP) layer, which predicts several tokens at once for faster decoding, enabled through speculative decoding with vLLM and SGLang. A separate Hy3-FP8 checkpoint is also available, which lowers the memory footprint for cheaper serving. The research team published scores across coding, agents, and STEM, including 78.0 on SWE-Bench Verified, 57.9 on SWE-Bench Pro, and 75.8 on SWE-Bench Multilingual.

On STEM and reasoning, Hy3 reports 90.4 on GPQA Diamond and 72.0 on USAMO 2026. The research team ran a blind test with 270 experts, collecting 312 valid comparisons on real workflows, where Hy3 scored 2.67 out of 4, ahead of GLM-5.1 at 2.51. The team focused on production reliability, addressing three failure modes backed by internal numbers.

Hy3 exposes an OpenAI-compatible API, deployable with vLLM or SGLang. The model is built around agent-style, long-context work, with a focus on size, not just score. Hy3 trades some coding accuracy for a far smaller active footprint, which matters when self-hosting and paying for GPUs.

Serving requires real memory, and the research team recommends 8 GPUs. Tencent provides a complete finetuning pipeline for Hy3 and points to its AngelSlim toolkit for compression. Why this matters: The release of Hy3 has significant implications for the AI industry, particularly in the areas of reasoning, agentic workflows, and long-context tasks.

With its sparse MoE architecture and MTP layer, Hy3 offers a more efficient and scalable solution for developers and businesses. The model's performance on coding, STEM, and reasoning tasks demonstrates its potential to improve productivity and accuracy. As the AI landscape continues to evolve, Hy3's open-source availability and compatibility with popular frameworks like vLLM and SGLang make it an attractive option for those looking to leverage MoE models in their applications.

However, questions remain about the model's limitations and potential biases, and further research is needed to fully explore its capabilities and implications.