Author's Description
MAI-DS-R1 is a post-trained variant of DeepSeek-R1 developed by the Microsoft AI team to improve the model’s responsiveness on previously blocked topics while enhancing its safety profile. Built on top of DeepSeek-R1’s reasoning foundation, it integrates 110k examples from the Tulu-3 SFT dataset and 350k internally curated multilingual safety-alignment samples. The model retains strong reasoning, coding, and problem-solving capabilities, while unblocking a wide range of prompts previously restricted in R1. MAI-DS-R1 demonstrates improved performance on harm mitigation benchmarks and maintains competitive results across general reasoning tasks. It surpasses R1-1776 in satisfaction metrics for blocked queries and reduces leakage in harmful content categories. The model is based on a transformer MoE architecture and is suitable for general-purpose use cases, excluding high-stakes domains such as legal, medical, or autonomous systems.
Key Specifications
Supported Parameters
This model supports the following parameters:
Features
This model supports the following features:
Performance Summary
Microsoft's MAI-DS-R1, a post-trained variant of DeepSeek-R1, demonstrates a balanced performance profile with a strong emphasis on safety and reasoning. The model exhibits moderate speed performance, ranking in the 31st percentile, and offers moderate pricing, placing it in the 38th percentile. A standout feature is its exceptional reliability, boasting a 99% success rate across benchmarks, indicating minimal technical failures. MAI-DS-R1 excels in several key areas. It achieves perfect accuracy in General Knowledge, Ethics, and one of the Instruction Following benchmarks, with the latter also being a top performer in speed. Its Coding and Reasoning capabilities are highly competitive, scoring 95% and 98% accuracy respectively, placing it in the 96th and 94th percentiles. The model also performs well in Email Classification (99% accuracy). While its Hallucinations (Baseline) accuracy is 88.9%, placing it in the 34th percentile, this indicates some room for improvement in acknowledging uncertainty. Mathematics performance is solid at 90% accuracy. Overall, MAI-DS-R1 successfully unblocks previously restricted topics while maintaining strong core AI capabilities, making it suitable for general-purpose use cases.
Model Pricing
Current Pricing
Feature | Price (per 1M tokens) |
---|---|
Prompt | $0.25 |
Completion | $1 |
Price History
Available Endpoints
Provider | Endpoint Name | Context Length | Pricing (Input) | Pricing (Output) |
---|---|---|---|---|
Chutes
|
Chutes | microsoft/mai-ds-r1 | 163K | $0.25 / 1M tokens | $1 / 1M tokens |
Benchmark Results
Benchmark | Category | Reasoning | Strategy | Free | Executions | Accuracy | Cost | Duration |
---|
Other Models by microsoft
|
Released | Params | Context |
|
Speed | Ability | Cost |
---|---|---|---|---|---|---|---|
Microsoft: Phi 4 Reasoning Plus | May 01, 2025 | ~14B | 32K |
Text input
Text output
|
★ | ★★★ | $$$$ |
Microsoft: Phi 4 Multimodal Instruct | Mar 07, 2025 | ~5.6B | 131K |
Text input
Image input
Text output
|
★★ | ★★ | $$ |
Microsoft: Phi 4 | Jan 09, 2025 | ~14B | 16K |
Text input
Text output
|
★★★★ | ★★★★ | $$ |
Microsoft: Phi-3.5 Mini 128K Instruct | Aug 20, 2024 | ~3.8B | 128K |
Text input
Text output
|
★ | ★★ | $$ |
Microsoft: Phi-3 Mini 128K Instruct | May 25, 2024 | ~3.8B | 128K |
Text input
Text output
|
★★★ | ★★ | $$ |
Microsoft: Phi-3 Medium 128K Instruct | May 23, 2024 | ~14B | 128K |
Text input
Text output
|
★★ | ★ | $$$$ |
WizardLM-2 8x22B | Apr 15, 2024 | 22B | 65K |
Text input
Text output
|
★★★ | ★★ | $$$ |