Tencent dropped Hy4 Preview today.
It immediately won ISS test. Same prompt: build a 3D simulation of the ISS in orbit. Hy4 Preview: 1h 52m · $1.30 Qwen 3.8 Flash-Next: 2h 03m · $0.30 GLM 5.3: 1h 46m · $3.20 DeepSeek V4 Flash Vision: 57m · $1.20 Hy4 was the surprise winner for
Even though the main models of Chinese AI are "giant models", their winning paths are different.
The Hy4 preview released by Tencent has a total of 770 billion parameters. Instead of moving all of that every time, it's 49 billion that moves every single token.
CARNICE-V3-27B just made local AI agents way more interesting.
But the benchmark claim isn’t even the part I’d pay attention to. What Carnice actually is: → 27.8B parameters, built on Qwen 3.8-27B → Apache 2.0, open-source, commercial use allowed → Fine-tuned specifically
You have reached the end of the archive
All of qwen38