Tencent Open-Sources Hy3 at 74.4%

Key Takeaways

• Hy3 preview scores 74.4% on SWE‑bench, closing gap to top models
• Model activates only 21 B parameters per token, cutting inference cost 40%
• API at RMB 1.2 per million tokens (~$0.18), 10‑15× cheaper than rivals
• Tencent rebuilt its AI stack in 90 days, launching Hy3 in apps
• MoE routing with 192 experts per layer enables performance on GPU clusters

Pulse Analysis

The open‑source large‑language‑model market has entered a new phase of competition as Tencent released Hy3, a 295‑billion‑parameter mixture‑of‑experts model that activates just 21 billion parameters per token. Scoring 74.4% on the SWE‑bench Verified benchmark, Hy3 narrows the gap with industry leaders like Claude Opus (80.8%) while surpassing other Chinese offerings such as GLM‑5 and Kimi‑K2.5. The rapid 90‑day rebuild of Tencent’s AI stack, from data pipelines to reinforcement‑learning loops, underscores a strategic shift toward cost‑effective, production‑ready agents rather than pure parameter scaling.

Hy3’s architecture leverages a sparse MoE design with 192 routed experts per layer and top‑k routing, allowing the model to deliver frontier‑grade performance with substantially lower compute per token. In internal tests, first‑token latency fell 54% and end‑to‑end latency dropped 47%, while inference costs were cut by roughly 40% compared with dense counterparts. Priced at RMB 1.2 per million input tokens—about $0.18—Hy3 offers a cost‑to‑capability ratio that is 10‑15 times cheaper than comparable closed‑source APIs. The model is available through OpenRouter with free two‑week access, and can be served on a modest four‑GPU A100 cluster using vLLM or SGLang, making high‑performance LLMs accessible to midsize enterprises.

The release sends a clear signal to the broader AI ecosystem: open‑source models can now compete on both performance and economics. Tencent’s swift stack overhaul and aggressive pricing pressure incumbents such as Anthropic, OpenAI and Microsoft to accelerate their own cost‑reduction strategies. For enterprises, Hy3 expands the viable options for building agentic workflows in coding, search and productivity tools, potentially accelerating AI adoption across Chinese and global markets. As more firms experiment with MoE‑based open models, the balance of power in the LLM landscape may shift toward providers that can deliver frontier capabilities without prohibitive cloud spend.