Tencent releases Hy3, a 295B-parameter MoE model
Tencent released Hy3 on July 6 under Apache 2.0, publishing standard and FP8 weights for a 295-billion-parameter mixture-of-experts model that activates 21 billion parameters per token and supports a 256K context window. The company also published vLLM and SGLang deployment recipes; its benchmark, hallucination and product-reliability results remain vendor-reported pending broader independent evaluation.
Tencent released Hy3 on July 6, moving its 295-billion-parameter mixture-of-experts model from an April preview to an Apache 2.0 release with downloadable standard and FP8 weights. The official model card and repository identify Tencent's Hy team as the developer and provide deployment paths for vLLM and SGLang.
What Tencent shipped
Hy3 has 295 billion total parameters, with 21 billion active for each token and a separate 3.8-billion-parameter multi-token-prediction layer. Tencent lists 192 routed experts, top-eight routing and a 256K context window. The weights are available through Hugging Face and other model repositories, while the official documentation includes serving, fine-tuning, reinforcement-learning and quantization guidance.
The active-parameter count affects the compute used for each token, but it does not make the full model small. Tencent recommends eight H20-3e GPUs or other high-memory accelerators for an eight-GPU deployment. Teams evaluating Hy3 still need to account for weight storage, interconnect bandwidth, KV-cache growth at long context lengths and the operational maturity of the serving stack.
Performance evidence is mostly vendor-reported
Tencent says a blind evaluation involving 270 experts gave Hy3 a score of 2.67 out of 4, compared with 2.51 for GLM-5.1. It also reports that an internal hallucination measure fell from 12.5% to 5.4% and that multi-turn issue rates fell from 17.4% to 7.9%. Those figures describe Tencent's own evaluations and product feedback, not an independent reproduction.
Independent coverage from Caixin and The Decoder confirms the July 6 release, the model's 295B/21B architecture and its Apache 2.0 availability. It does not independently validate Tencent's benchmark or reliability claims.
What practitioners should test
For production comparisons, the useful questions are workload-specific: throughput at realistic prompt lengths, memory use, tool-call stability, quantization quality and behavior across different agent scaffolds. Hy3's permissive license and official deployment recipes make those tests practical, but the model's scale still puts self-hosting beyond a typical single-GPU setup. The release is therefore a credible open-weight option to evaluate, not proof that Tencent's reported gains will transfer unchanged to every environment.
Key Points
- 1Tencent released Hy3 and Hy3-FP8 under Apache 2.0 on July 6 with official model and deployment artifacts.
- 2The MoE model has 295B total parameters, 21B active parameters per token and a 256K context window, but still requires substantial multi-GPU infrastructure.
- 3Tencent's benchmark and reliability improvements are vendor-reported and need independent, workload-specific validation.
Scoring Rationale
Hy3 is a significant open-weight model release because it combines a 295B MoE architecture, 21B active-parameter path, 256K context and Apache 2.0 artifacts. Its practical impact remains bounded by substantial hosting requirements and the need for independent validation of Tencent's performance and reliability claims.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
