Word counting matches MAI-Thinking-1
Two AI models that appear to be unreleased Microsoft models surfaced in quick succession on Oct. 24 (Oct. 25 Japan time) with their names withheld. @aitrackerbot, an X account that tracks the appearance of models, posted its analysis.
One is "iceman-contributor-free," which appeared in "OpenCode Go," a paid service of the programming tool "OpenCode." According to @aitrackerbot, the number of units (tokens) it splits the same prompt into matched Microsoft's "MAI-Thinking-1" exactly. When asked, the model reportedly identified itself as MAI. It handles text only, supports reasoning and tool calling, and has a context window of 262,000 tokens.
MAI-Thinking-1 is Microsoft's first reasoning model, released at the company's Build developer conference in June. It has 35 billion parameters, and Microsoft said it outperformed Anthropic's "Sonnet 4.6" on programming evaluations and matched "Opus 4.6."
Image model's pixel characteristics match MAI-Image
The other is "pepitas," an image generation model that appeared on Arena, a site that compares AI models anonymously. When asked to draw a self-portrait, it drew itself wearing an OpenAI T-shirt, but when @aitrackerbot classified it by pixel characteristics, it matched Microsoft's "MAI-Image." It did not match image models from OpenAI or China's Tencent.
Microsoft released the image model "MAI-Image-2.5" in May, and it ranks third on Arena. It is not known whether pepitas is its next version.
Preparing a string of in-house models?
Microsoft is expanding its lineup of in-house "MAI" models to reduce its reliance on OpenAI. On the 20th, the unreleased "MAI-Code-2" was reported to have appeared in GitHub Copilot's list of preview models. Both analyses were made by an outside account, and ORIENS has not independently verified them.
