Tencent releases Hy3, a 295B open Mixture-of-Experts model

Category: AI
Status: confirmed
Sources: Tencent announcement, Hugging Face weights, HN discussion
Summary: Tencent Hunyuan released Hy3 on 2026-07-06 under Apache 2.0, a Mixture-of-Experts model with 295B total parameters and about 21B active per token and a 256K context window, positioned for agentic and long-context coding work. Tencent states the release follows a late-April preview and feedback from more than 50 internal product teams, and reports it rivals open models with 2 to 5 times the parameter count. Weights are on Hugging Face and ModelScope, and the model is free on OpenRouter until 2026-07-21.
Comments: HN commenters report the model performs well for its size, placing it near Sonnet 5 on some tasks and calling it a plausible local-inference option given the small active-parameter count. Some rank it below GLM 5.2 and GPT-5.5.
Why it matters: Another Apache-2.0 open-weight model with a low active-parameter cost adds pressure on both closed frontier pricing and larger open models.
Follow-up: Watch for independent benchmarks against GLM 5.2 and DeepSeek-V4, and local-inference reports at the 21B active size.