Atlas

Models

← All models

MAI-DS-R1

Microsoft Research · 2025-04-17 · 671B parameters

MAI-DS-R1 is Microsoft’s open-weight post-trained DeepSeek-R1 variant, tuned on about 110,000 Tulu 3 safety and non-compliance examples plus approximately 350,000 internally developed multilingual blocked-topic examples to improve responsiveness and its risk profile while preserving reasoning capabilities. Its released BF16 safetensors checkpoint contains exactly 671,026,419,200 parameters and is licensed under MIT.

Benchmark scores

BenchmarkScore
SnakeBench Average Apples2.0
SnakeBench Best Apples13.0
SnakeBench Rating16.2
SnakeBench Total Apples32.0
SnakeBench Win Rate31.3
Loading Atlas data…