Microsoft unveiled a family of seven on-premises AI models at its annual Build developer conference in San Francisco’; This is the clearest sign that the tech giant is taking action to reduce its dependence on AI companies, with billions of dollars pouring in.
After tuning its models for consulting firm McKinsey, Microsoft AI general manager Mustafa Suleiman said the company outperformed OpenAI’'s GPT-5.5’ in terms of quality and predicted ten times better cost efficiency based on public pricing data scaled to model sizes.
At the conference, CEO Satya Nadella, “, said, "We believe it's time for every company to move from consuming a border model to fully participating in the border."
The title version is MAI-Thinking-1’, Microsoft’'s first reasoning model; trained from the ground up with clean, commercially licensed data without being distilled from third-party systems.
It is a medium-sized model with 35 billion active parameters and a 256.000 coin context window, designed for complex multi-step instructions, long context reasoning, and code generation.
In addition, the company has launched MAI-Code-1-Flash’, an encoding model that converts text descriptions to source code for apps and websites, and is now available on GitHub Copilot and Visual Studio Code’.
By running its own models on Azure infrastructure, Microsoft can avoid the fees it already pays third-party providers and pass the savings on to developers.
According to Microsoft’, in blind reviews conducted by independent human rating partner Surge, MAI-Thinking-1 was preferred to Anthropic’in Claude Sonnet 4.6’, and the company said it matched Claude Opus 4.6 in its coding criteria.
Category: Quantum leaps
Also at Build’, Microsoft announced that the Majorana 2 quantum chip is 1,000 times more reliable than its predecessor; This is a milestone that the company says has come a striking distance to a commercially useful quantum computer.
Quantum qubits, the basic unit of quantum computing, are, as everyone knows, fragile: Even the slightest temperature changes or vibrations can lead them astray. The Majorana 2 chip handles this directly.
The qubits in the new chip survive an average of 20 seconds compared to the milliseconds in the original; This is a development that the company likens to upgrading a phone that needs to be charged daily to a phone that can last several years.
Zulfi Alam, corporate vice president of Microsoft Quantum’, said that in “2029’, we will have a quantum machine that can solve commercially viable, reasonable problems. While the chip currently has only 12 qubits, a handy machine requires millions of qubits.
Microsoft’'s approach focuses on topological qubits based on the properties of the semiparticle, first theorized by Italian physicist Ettore Majorana in the 1930s.
The path was not smooth – company had to withdraw a 2018 paper claiming evidence for the particle in the journal Nature – but the pressure continued and the second generation chip partially recovered from the first by replacing aluminum as a superconductor with lead.
The chip and the research supporting it have not yet undergone peer review, and some physicists have requested more information.
IPO race escalates
Microsoft’'s drive for model independence comes as companies allocate billions of dollars to prepare for blockbuster stock market launches.
Just days after raising $65 billion (59 billion euros) in the Series H funding round, which raised its valuation to $965 billion (877 billion euros), Anthropic, the artificial intelligence lab behind Claude’, secretly filed for an initial public offering on June 1, ’. OpenAI is also preparing its own secret file.
Microsoft has committed $13 billion (€11.8 billion) to OpenAI’, and Anthropic’e has invested up to $5 billion (€4.5 billion) while making models from both companies available through Azure.
