Code Generation on MultiPL-E MBPP
58.8ScoreKimi-K2
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Kimi-K2Model Variant=Base, # Shots=0-shot, # Activated Params=32B, # Total Params=1043B2026.02 | 58.8 | — | |
| Kimi-K2 Base# Shots=0-shot, # Activated Params=32B, # Total Params=1043B2026.01 | 58.8 | — | |
| Step 3.5 FlashModel Variant=Base, # Shots=0-shot, # Activated Params=11B, # Total Params=196B2026.02 | 58 | — | |
| MiMo-V2 FlashModel Variant=Base, # Shots=0-shot, # Activated Params=15B, # Total Params=309B2026.02 | 56.7 | — | |
| MiMo-V2-Flash Base# Shots=0-shot, # Activated Params=15B, # Total Params=309B2026.01 | 56.7 | — | |
| DeepSeek V3.1Model Variant=Base, # Shots=0-shot, # Activated Params=37B, # Total Params=671B2026.02 | 52.5 | — | |
| DeepSeek-V3.1 Base# Shots=0-shot, # Activated Params=37B, # Total Params=671B2026.01 | 52.5 | — | |
| DeepSeek V3.2Model Variant=Exp Base, # Shots=0-shot, # Activated Params=37B, # Total Params=671B2026.02 | 50.6 | — | |
| DeepSeek-V3.2 Exp Base# Shots=0-shot, # Activated Params=37B, # Total Params=671B2026.01 | 50.6 | — | |
| FullTraining Strategy=Full-Attention baseline2026.06 | — | 82.1 | |
| MSA-CPTTraining Strategy=sparse continued pretraining2026.06 | — | 81.1 | |
| MSA-PTTraining Strategy=from-scratch sparse pretraining2026.06 | — | 81.6 |