Hy3

About

Hy3 (tencent/Hy3, released July 2, 2026 under the Apache License 2.0) is the Tencent Hy Team's 295B-parameter Mixture-of-Experts model with 21B active parameters plus a 3.8B native MTP layer, introduced after the late-April Hy3 Preview launch with post-training scaled up using feedback from 50+ products. It outperforms similar-size models and rivals flagship open-source models with 2-5x parameters.

Architecture: 80 layers (first dense, then MoE) plus 1 native MTP layer, GQA attention (64 query heads, 8 KV heads, head dim 128) with QK-Norm, 192 routed experts with top-8 activation plus 1 shared expert (sigmoid router with expert bias), hidden size 4096, MoE intermediate size 1536, 256K context, vocabulary 120832, BF16. Hy3 shows solid gains in reasoning, agentic and long-context tasks, and in productivity scenarios (coding, office work, financial modeling, frontend design, game development). In a blind evaluation by 270 experts on real work tasks it scored 2.67/4, outperforming GLM-5.1 at 2.51/4; on SWE-Bench Verified, accuracy variance across scaffoldings (CodeBuddy, Cline, KiloCode) stays within 4%. Product-focused improvements include production-grade tool-call stability, anti-hallucination training (hallucination rate 12.5% -> 5.4%, commonsense errors 25.4% -> 12.7%), and complex multi-turn context retention (issue rate 17.4% -> 7.9%).

Training Data Predecessor of Hy4 preview; scores in the Hy4 preview model card benchmark appendix (re-evaluated by Tencent under the updated harness).