Hy3
Points and comments are a snapshot, not live.
Tencent's Hy3 model rivals larger open-source models with improved agent capabilities.
Tencent introduces Hy3, a 295B-parameter open-source model (Apache 2.0) that outperforms similar-size models and competes with flagship models 2-5x its size. Post-training improvements include stronger agent capabilities, with Hy3 scoring 2.67/4 in blind expert evaluation vs GLM-5.1's 2.51/4. Internal WorkBuddy tests showed task success rate rising from 72% to 90% and 34% faster completion. Hallucination rate dropped from 12.5% to 5.4%. API pricing: 1 RMB input, 4 RMB output per 1M tokens.
What commenters are saying
Commenters praise Hy3's performance relative to size and price, comparing it favorably to GPT-5.4-mini and GLM-5.2. Some note it is not at GPT-5.5 or GLM-5.2 tier. A user reports Hy3 stays on track more reliably than DeepSeek V4 Flash in agentic coding tasks, though DSV4 Flash benefits from architectural KV cache efficiency for long contexts. Concerns about benchmark contamination are raised. Hy3 is seen as a strong local model candidate for high-end hardware (e.g., 2 DGX Sparks), with quantization resilience unknown.