Tencent’s Hy team has released Hy3, a 295-billion-parameter Mixture-of-Experts (MoE) model that activates only 21 billion parameters per token while supporting a 256,000-token context window. The weights are available under the Apache License 2.0, and the model is specifically designed for reasoning, agentic workflows, and long-context tasks. With 192 experts and top-8 routing, Hy3 keeps computational costs low while delivering competitive performance across coding, STEM, and agent benchmarks. An FP8 quantized checkpoint, Hy3-FP8, is also available for reduced memory footprint during serving.<\/p>\n
Hy3 is a sparse MoE model with 192 experts and top-8 routing \u2014 only eight experts fire per token, which keeps inference efficient. The architecture includes a Multi-Token Prediction (MTP) layer that predicts several tokens at once, enabling faster decoding through speculative decoding in both vLLM and SGLang. The model has 80 layers (excluding the MTP layer), 64 attention heads with Grouped Query Attention using 8 KV<\/a> heads and a head dimension of 128, a hidden size of 4096, an intermediate size of 13312, and a vocabulary of 120,832 tokens. The context window spans 256K tokens, and the model supports BF16 precision.<\/p>\n The MTP layer itself contains 3.8 billion parameters. Total model size is 295B parameters, with only 21B activated per token, which is the core efficiency advantage of the MoE design.<\/p>\n Tencent’s research team published scores across coding, agents, and STEM domains. On coding benchmarks, Hy3 achieves 78.0 on SWE-Bench Verified, 57.9 on SWE-Bench Pro, and 75.8 on SWE-Bench Multilingual. Terminal-Bench 2.1 lands at 71.7, and DeepSWE at 28.0. On STEM and reasoning, the model reports 90.4 on GPQA Diamond, 72.0 on USAMO 2026, 90.0 on IMOAnswerBench, and 53.2 on HLE (with tools).<\/p>\n In a blind test with 270 experts collecting 312 valid comparisons on real workflows, Hy3 scored 2.67 out of 4, ahead of GLM-5.1 at 2.51. The largest advantage appeared in frontend development, CI\/CD, and data and storage tasks.<\/p>\n Tencent’s team focused three specific areas of production reliability. First, tool calling and output formatting received improvements that reduced invalid calls causing infinite loops. Accuracy<\/a> variance across CodeBuddy, Cline, and KiloCode scaffoldings stays within 4% on SWE-Bench Verified. Second, anti-hallucination training reduced the hallucination rate from 12.5% to 5.4% and commonsense error rates from 25.4% to 12.7% in internal evaluations. Third, multi-turn intent tracking improved through joint supervised fine-tuning and reinforcement learning, dropping the internal issue rate from 17.4% to 7.9% and raising MRCR long-dialogue benchmark scores from 42.9% to 75.1%.<\/p>\n Hy3 exposes an OpenAI-compatible API. You deploy it with vLLM or SGLang, then call the endpoint. A single flag, reasoning_effort<\/strong>, controls how much the model thinks before responding. Use no_think<\/strong> for direct answers, low<\/strong> for light reasoning, and high<\/strong> for deep chain-of-thought on math, coding, or multi-step tasks. The recommended settings are temperature 0.9 and top_p 1.0. You can also try Hy3 without local hardware \u2014 OpenRouter lists a tencent\/hy3:free route at no cost, with the free tier scheduled to end on July 21, 2026.<\/p>\nBenchmark and Performance<\/h2>\n
Reliability and Production Behavior<\/h2>\n
How to Call Hy3<\/h2>\n
Hy3 vs GLM-5.2: Size Versus Raw Performance<\/h2>\n