Recently, Microsoft released the decision model Microsoft-Decision-1, the first version of which is built on Alibaba Cloud's open-source model Qwen3.5-9B, primarily for text classification and structured decision-making tasks. The model is now available via Microsoft Foundry and will be integrated with OpenRouter in the future. Microsoft stated that future versions will shift to using proprietary models and OpenAI models as their foundation.
Unlike large language models primarily used for generating text or handling complex reasoning tasks, decision models typically output a limited range of results for specific questions and provide probability scores for different options. These models generate faster, have relatively lower usage costs, and their outputs are more easily called directly by software systems, making them suitable for enterprise scenarios that do not require open-ended text. However, decision models can still produce classification or judgment errors.
Microsoft-Decision-1 performs excellently in benchmark tests
Test results released by Microsoft show that Microsoft-Decision-1's speed in latency tests is approximately 2.5 times that of H2O-Lightning-4B and 2.8 times that of Jev. Across 36 benchmark tests, the model achieved an accuracy of 83.5%; a confidence score of 92.2%, ranking second only to Quyet-1.0-Large. Microsoft also stated that in text classification tasks, its cost is more than 20 times lower than OpenAI's GPT-6 Sol.
Currently, input is priced at US$0.04 per 1 million tokens (approximately HK$0), while output is free.
Recently, decision models have attracted attention from multiple AI companies. Following TypeSafe AI's release of Jev, companies including OpenAI, Cloudflare, Liquid AI, Perplexity, Snowflake, and H2O.ai have also launched related models or APIs. Over 100 such models have now entered the market, among which H2O-Lightning-4B is likewise developed based on the Qwen3.5 series of models.