Tencent Releases Hunyuan Hy3: 295B Open-Weight MoE Model Under Apache 2.0
Tencent
Tencent officially launched Hunyuan Hy3 on July 6–7 — a 295B Mixture-of-Experts model with only 21B active parameters per forward pass, a 256K-token context window, and full Apache 2.0 weights. A hybrid fast/slow-thinking system targeting agentic coding, complex reasoning, and instruction-following. Tencent claims intelligence comparable to flagship models 2–5x its active parameter scale. Weights are on HuggingFace (`tencent/Hy3`) and ModelScope in BF16 and FP8. Free API access via OpenRouter until July 21.
Why it matters
One of the largest open-weight MoE models released by a Chinese lab under fully permissive licensing (Apache 2.0, no geographic carve-outs), directly competing with DeepSeek V4 and GLM-5.2 at a fraction of inference cost. The 295B/21B active parameter ratio runs like a much smaller model while claiming frontier-class reasoning scores.
Importance: 4/5
Major open-weight MoE release from Tencent under Apache 2.0; direct competitor to DeepSeek V4 with free trial access.