Explore Yuan 3.0 Ultra, YuanLab AI's powerful new model. Discover its advanced features for intelligence and efficiency. Learn more today!
Research on expert load distribution shows two main parts during training:
- Early Stage: Expert loads change a lot because they start randomly.
- Stable Stage: Expert loads settle down. The order of experts for processing data stays mostly the same.
- Small Load Rule (⍺): This helps experts with much less work than average.
- Total Load Rule (β): This finds experts that do the least work overall.
Faster Hardware and Better Expert Setup
| Method | TFLOPS per GPU |
| Base Model (1515B) | 62.14 |
| DeepSeek-V3 Aux Loss | 80.82 |
| Yuan3.0 Ultra (LAEP) | 92.60 |
- Model Pruning: Helped make it 32.4% more efficient.
- Expert Rearrangement: Helped make it 15.9% more efficient.
Less Overthinking with New RIRM Method
- rmin=0: Best for quick, direct answers.
- rmax=3: The highest number of checks allowed.
How Yuan 3.0 Ultra Does on Business Tests
| Test | What it Tests | Yuan3.0 Ultra Score | Top Competitor Score |
| Docmatix | Multimodal RAG | 67.4% | 48.4% (GPT-5.2) |
| ChatRAG | Text Search (Avg) | 68.2% | 53.6% (Kimi K2.5) |
| MMTab | Table Questions | 62.3% | 66.2% (Kimi K2.5) |
| SummEval | Summaries | 62.8% | 49.9% (Claude Opus 4.6) |
| Spider 1.0 | Text-to-SQL | 83.9% | 82.7% (Kimi K2.5) |
| BFCL V3 | Using Tools | 67.8% | 78.8% (Gemini 3.1 Pro) |
#Yuan30Ultra #AIModel #MultimodalAI #MoEModel #AIEfficiency #LargeLanguageModels #ArtificialIntelligence
