Tencent’s Hy3 draws developer scrutiny on OpenRouter

Tencent’s Hy3 is drawing attention as a new Apache 2.0 licensed model from China, with availability through OpenRouter and a free option from Novita until July 21st. Early users highlighted its stronger results versus the preview version, including improvements on coding-related benchmarks such as DeepSWE, while cautioning that benchmark gains may not translate cleanly into real workloads.

Discussion centered on how Hy3 compares with DeepSeek v4 Flash, GLM-5.2, Qwen 3.6 27B and MiMo models. One quoted specification says Hy3 has 295B parameters in total and recommends serving it on 8 GPUs, putting it outside typical consumer local setups. Developers also debated whether its pricing and benchmark profile make it competitive with DeepSeek’s cheaper Flash tier and more capable Pro variants.

Hands-on reports were mixed. Some users praised Hy3 for following instructions, staying on track in coding sessions and offering strong capability for its size and price. Others reported slow responses, HTTP errors, heavy rate limiting and weaker performance than smaller local models in specific tasks. Local inference comparisons also raised concerns about KV cache efficiency, with one user contrasting DeepSeek v4 Flash support for 3M tokens of KV cache against Hy3 at roughly ~130K tokens under their setup.