Open models
Open-weight language models compared on capability, licensing and deployment.
MMLUGPQA DiamondHumanEval MATH
Llama 3.3 70BDense 70B, 128K8650.588.477
Qwen 2.5 72BDense 72B, 128K86.149.888.276.8
DeepSeek V3MoE 671B, 128K87.552.189.380.2
Falcon 3 10BDense 10B, 32K68.531.265.424.7
Recommended in the report
Findings
- 01
The distinction between true open-source software, as defined by the Open Source Initiative, and open-weights releases carries significant implications for vendor lock-in, legal indemnification, and algorithmic auditability.
p. 2