Ranked by MMLU score. Only includes models with open weights (Apache 2.0, MIT, Llama Community License, etc.).
Updated April 2026 · Source: AI Flash Report model database
| # | Model | MMLU | Parameters | Context | License |
|---|---|---|---|---|---|
| #1 |
DeepSeek V3.2
DeepSeek
|
90.1% | 671B MoE | 1M tokens | DeepSeek License (open weights, commercial OK) |
| #2 |
DeepSeek-V3
DeepSeek
|
88.5% | 671B | 128K tokens | DeepSeek License (open weights) |
| #3 |
Grok-1
xAI
|
73.0% | ~314B | 8K tokens | Apache 2.0 |
| #4 |
Kimi K2
Moonshot AI
|
91.3% | 1.04T MoE | 2M tokens | Modified MIT (open weights) |
| #5 |
Llama 2 70B
Meta
|
68.9% | 70B | 4K tokens | Llama 2 Community License |
| #6 |
GLM-5
Zhipu AI
|
88.7% | 744B | 200K tokens | Proprietary (open weights for non-frontier sizes) |