Microsoft releases the decision model Microsoft-Decision-1.1, with 2x faster speed and multimodal support
Microsoft has released a new-generation decision model, Microsoft-Decision-1.1, which scored 68 on the Artificial Analysis decision index leaderboard, 6 points higher than the first decision model, with more than 2x faster computation and lower cost, and added support for image input. The model is developed based on Gemma 4 26B A4B, uses a custom inference engine, and is priced at $0.042 per million input tokens in Foundry, with output free of charge. Microsoft also released the coding models MAI Code 2, MAI Code 2 Flash, etc., and integrated the decision model into Copilot for request routing, processing 2.2x faster and at about 1/4 the cost of large language models.
Why read this
AI score 66
Microsoft's decision model's trade-offs in latency, cost, and calibration, as well as how it is bound to Copilot and Cowork, can help judge the productization path for specialized small models.
Topics
Body · AI translation
Zhihui Dongxi news on November 18: Today, Microsoft released its new-generation decision model, Microsoft-Decision-1.1. On the Artificial Analysis decision index leaderboard, Microsoft-Decision-1.1 scored 68 points, ranking 8th, 6 points higher than Microsoft's first decision model, and ahead of Jev 1.14, Google Gemini Decision 1, and OpenAI's GPT-6 Luna version decision model.
On October 10, Microsoft launched its first in-house decision model, Microsoft-Decision-1. Microsoft says that compared with the previous-generation model, Microsoft-Decision-1.1 has stronger intelligence and more precise calibration capabilities, runs more than 2 times faster, and costs less.
In addition to the decision model, at the Ignite conference Microsoft also released the programming models MAI Code 2 and MAI Code 2 Flash, built specifically for enterprises' everyday coding tasks, as well as new models Code 2 and Code 2 Flash built specifically for programming and cybersecurity, among others. Among them, in three coding benchmark tests measuring everyday high-volume routine coding tasks, the performance of MAI Code 2 Flash is comparable to GPT-6 Luna and Claude Haiku 5.5, at only 65% and 20% of their cost.
The new-generation decision model adds multimodal support and can accept images. On pricing, in Microsoft Foundry, it costs $0.042 per million input tokens (about RMB 0.28), and output is free.
Microsoft says its decision models are designed to make calibrated decisions under uncertainty across a variety of real-world software use cases. They choose from allowed actions within constraints set by developers and output calibrated confidence scores, so that applications can defer decisions when uncertainty is high.
In addition, because these models focus on decision tasks, they are significantly faster and cheaper than general-purpose large language models when handling decision-related tasks.
Microsoft carried out joint innovation in latency, cost, intelligence, and calibration:
Latency: Decision models need to run in real time, even when handling complex requests with long inputs and hundreds of questions. Unlike traditional large language model inference engines optimized for token-by-token text generation, Microsoft-Decision-1.1 uses a custom inference engine optimized for the computational needs of decision models.
Cost: Microsoft's custom inference engine achieves low-cost optimization by efficiently distributing requests across GPU clusters, rather than pursuing only low latency. It is developed based on the mid-sized mixture-of-experts model Gemma 4 26B A4B. Gemma 4 26B A4B has enough active parameters to provide the intelligence needed for decision-making, while being efficient enough to run on a variety of flexible hardware configurations such as A100, H100/200, and GB200/300.
Intelligence: Decision models need to remain small enough to run in all scenarios, generalize well to new or unseen tasks, and be specially trained to make decisions. To achieve this, Microsoft combines supervised fine-tuning with reinforcement learning, using hundreds of synthetic and licensed datasets covering tens of thousands of task types from numerous industries, to train the model to make judgments rather than memorize answers, enabling the model to distinguish between highly similar options; identify uncertainty; and clearly determine when to answer and when to refuse to answer.
Calibration: Microsoft trained the model to explicitly understand its own cognitive boundaries, give consistent answers in similar scenarios, accurately report the confidence of its answers, split confidence when multiple valid answers exist, and refuse to answer when information is incomplete. Developers can set their own thresholds for decisions, flags, or escalations.
In addition to third-party testing, Microsoft also evaluated Microsoft-Decision-1.1 on 102 benchmarks, covering more than 385,000 questions not encountered during training, with test tasks including routing, ranking, long context, multilingual, out-of-distribution tasks, reasoning, safety, and many other areas.
Microsoft-Decision-1.1 achieved the highest average accuracy among 8 decision models and leads in 7 of 9 capability categories.
In just over five weeks since launch, more than 4,000 customers have used Microsoft's decision models to complete various tasks, including model routing, customer support triage, business process automation, labeling, and classification. Early adopters include Epic, Equinor, Atos Group, and Raiffeisen Bank International of Austria.
At the same time, Microsoft is integrating the specialized intelligence of Microsoft-Decision-1.1 into various products to make fast and efficient decisions, while leaving larger models for more complex tasks.
In Copilot, Microsoft-Decision-1.1 routes user requests to the appropriate interaction scenario, determining whether Copilot or Cowork is best suited to handle them. The accuracy of this decision is comparable to that of large language models, while processing is 2.2 times faster and costs about 1/4 as much.
In Cowork, the model powers Auto, dynamically adjusting the amount of reasoning effort Cowork invests in each task.
Conclusion: Microsoft's models are deeply tied to its own products
and may take a differentiated route
Unlike other leading model vendors, Microsoft's difference lies in the deep integration of its models with its own products such as GitHub Copilot, Foundry, and Copilot workflows, with model capabilities embedded directly into existing office and development entry points rather than selling APIs separately.
This means enterprises can plug a complete solution directly into existing processes, with clear compliance boundaries and lower onboarding costs.
Source: Microsoft
This article comes from the WeChat public account "Zhihui Dongxi", author: Cheng Qian, editor: Li Shuiqing, authorized for publication by 36Kr.
AI Trends