Chinese AI Labs Launch Trillion-Parameter Open Model Offensive
모델 출시 | Fri Jul 17 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 3 sources
Moonshot's Kimi K3 and Ring-Zero 1T represent frontier challenges from ultra-large open models originating in China.
Analysis
[Moonshot AI] unveiled Kimi K3 open frontier model [2]
- World's first open 3T-class model at 2.8 trillion parameters
- 1M token context window and native vision support
- Adopts Kimi Delta Attention and Attention Residuals architecture
- Full model weights to be released by July 27
[Kimi K3] expanded MoE sparsity and improved scaling efficiency [2]
- Stable LatentMoE framework activating 16 of 896 experts
- Approximately 2.5x scaling efficiency improvement over Kimi K2
- Specialized for engineering tasks like long-horizon coding and GPU kernel optimization
- Supports game development
- frontend
- and CAD using screenshots and visual information
[Moonshot AI] pursued funding at $31.5B valuation [1]
- China's largest open-weight model with 2-3 trillion parameters
- Aims to narrow the gap with closed-source frontier models
- Raising funds again after $2B round at $20B valuation in May
- Emerging as open-source alternative alongside DeepSeek and Z.ai
[Ring-2.5-1T-Zero] published paper on 1 trillion parameter Zero RL reasoning model [3]
- Zero RL training with verifiable rewards without human-annotated data
- Stabilization techniques including clipped importance sampling and mixed-precision control
- Emergent cognitive behaviors such as self-verification and parallel reasoning
- Achieved competitive performance on 7 mathematics benchmarks