Recently, Tencent officially released the full weights for Hy3. This is a massive 295-billion-parameter Mixture-of-Experts (MoE) model. Importantly, Tencent published the Hy3 295B MoE model under the permissive Apache License 2.0. This launch serves as a direct follow-up to the initial Hy3-preview released earlier this year on April 23. You can download the official artifacts right now across Hugging Face, ModelScope, GitCode, and CNB.
Making Massive Models Accessible
Tencent built Hy3 to change the hardware economics of deploying massive models. The system features 295 billion total parameters. However, the MoE design utilizes 192 experts with top-8 routing. Consequently, the model activates only 21 billion parameters per token during inference.
This active-parameter path drastically lowers the barrier to entry. Mid-tier GPU clusters can handle the inference cost. As a result, organizations can access frontier-style reasoning without paying for hyperscaler infrastructure. Nevertheless, hosting the full 295B footprint still requires large multi-GPU hardware.
Furthermore, Tencent optimized this release for practical deployment. The package includes standard BF16 weights alongside FP8 variants. Additionally, the architecture incorporates 3.8 billion MTP layer parameters. Tencent also provided documented deployment recipes for vLLM and SGLang with speculative decoding enabled. Therefore, engineering teams can immediately test real throughput instead of just reading benchmark tables.
Tencent Optimized Hy3 295B MoE for STEM & Backend Tasks
Tencent explicitly designed Hy3 for utility rather than chatbot polish. Specifically, the model targets complex STEM reasoning and backend coding tasks. This specialization directly benefits research teams working in computational biology, materials science, and engineering design.
According to the official model card, the April preview achieved a 79.26 MMMLU 5-shot score. Tencent also claimed strong performance on the Spring 2026 Tsinghua Qiuzhen College Math PhD qualifying exam. However, the company did not publish an exact score for that specific test.
Meanwhile, the new July release features a massive 256K context window. Tencent claims this updated version delivers stronger post-training results. Specifically, they report gains in agent reliability, long-context handling, and product feedback tests.
Waiting on Independent Verification
Despite the impressive technical specifications, the verification gap remains real. Right now, all benchmark figures are self-reported vendor claims. Independent evaluators, such as Epoch AI, have not yet verified the data. The industry must wait to see if the high MMMLU scores actually translate to real-world task accuracy.
Consequently, production teams should use the official model card as a starting point, not a guarantee. Adopters need to evaluate latency under realistic 256K context lengths. Furthermore, engineers must test tool-call stability and memory footprints across non-Tencent enterprise agent scaffolds.
Finally, the open-weight MoE space is moving incredibly fast this year. Hy3 sits in meaningful company alongside DeepSeek V4, GLM, and Qwen. However, without broader cross-model evaluations, the industry cannot definitively verify any “largest open-source MoE” claims. For now, the weights are public, and practitioners can run their own evaluations to find out exactly what Hy3 can do.
