Tencent Hy4: Cheap, strong – and hallucinating

🔥 What happened Tencent released Hy4, its new flagship model. It uses a mixture-of-experts architecture with 770 billion parameters (49 billion active per token) and a one-million-token context window. New capabilities include image analysis and controlling computers via graphical interfaces and the command line. 💡 Why it matters Hy4 scores 35 points on the Artificial Analysis Intelligence Index – second among Chinese models and on par with Grok 4.7. On agentic coding (FrontierCode 2.0) it hits 31 percent, nearly matching Claude Fable 5.1 (32 percent). The kicker: at $0.78 per test task it's roughly ten times cheaper than Grok 4.7 – but its hallucination rate is a staggering 89 percent. ⚡ Our take Tencent built a price-performance monster that would rather make things up than admit ignorance. Anyone deploying Hy4 in production needs serious guardrails – otherwise the savings get expensive fast.