Overview
- Tencent has rolled out Hy4 Flash this week and is running it inside WorkBuddy and Yuanbao while offering API access on Tencent Cloud with published pricing and a 50% promotional discount through November 30.
- The model contains 224 billion total parameters with only 12 billion activated per inference, supports a 1,000,000‑token context window, accepts text and images, and uses Engram conditional memory and routing (LatentMoE) to reduce compute.
- Tencent reports decoding speeds above 150 tokens per second using low‑level optimizations and quantization to cut latency and run costs for routine tasks like document editing and slide generation.
- Vendor tests reported near‑equal user satisfaction to Anthropic’s Claude Haiku 5.5 while costing roughly 18 times less per task, but those results come from Tencent’s internal WorkBuddy study of 85 users and 259 tasks and are not independently validated.
- The release follows a broader Chinese trend of shipping smaller 'flash' variants after flagship launches; planned public weight releases and OpenRouter access could widen developer use but also raise deployment, safety, and dual‑use questions.