Tencent Launches Hy3 Large Language Model With 295 Billion Parameters

Tencent launched Hy3, the third generation of its flagship large language model, last week. The model uses a Mixture-of-Experts architecture with 295 billion total parameters and 21 billion activated parameters along with a 256K context window, Forbes reported. Independent evaluations placed Hy3 at 84.2 on BrowseComp and 79.1 on the public MCP-Atlas set, with a hallucination rate of 5.4 percent.

On coding benchmarks the model scored 78 percent on SWE-bench Verified, 71.7 on Terminal-Bench 2.1, and 28.0 on DeepSWE. A larger rival, GLM-5.2, recorded higher scores on the same coding tests at 84.2 percent, 81, and 46.2 respectively. GLM-5.2 contains 744 billion total parameters and activates roughly 40 billion per token.

Tencent said internal testing showed WorkBuddy task success rates rising from 72 percent to 90 percent and average execution time falling 34 percent. Hallucination rates in Yuanbao for long-document and search tasks dropped by more than half. The model is priced at approximately $0.18 per million input tokens and $0.59 per million output tokens through Tencent Cloud.

An FP8-quantized version runs on a single 8x H200 node using under 300 GB of memory. Daily token consumption has risen twenty-fold since the preview release, while the number of users actively selecting Hy3 inside WorkBuddy grew six times, according to the company. Forbes reported that Anthropic holds roughly 32 percent of the enterprise API market in 2026 compared with OpenAI at 25 percent.

Claude Code, Anthropic's coding agent, has reached an annualized revenue run rate of $2.5 billion.