This post argues that "Microsoft-Decision-1" names a role that different models will fill in turn, and that anyone who relies on the model's measured accuracy needs a record of which model is filling the role.
What Microsoft released
Microsoft put a model called Microsoft-Decision-1 into its Foundry cloud catalog on October 8 and 9. Satya Nadella announced it as "our new model for fast decision-making." It is a decision model: a model that does not write text. The caller supplies a question and a fixed list of answers, and the model returns a probability for each answer. Products use models like this to route requests, to sort them into categories, and to approve or block an agent's next action. This blog described the category on October 4, when Amazon and Cloudflare released their own.
Microsoft did not build this model from scratch. According to TestingCatalog, MarkTechPost, and OrcaRouter, all of which cite Microsoft's catalog page, Microsoft took Qwen3.5-9B, an open-weight model released by Alibaba, and trained it further for the scoring task. Further training of an existing model is called post-training. The model's knowledge and most of its parameters come from the base; post-training changes what it does with them.
The same sources report a second sentence from Microsoft's catalog, which OrcaRouter quotes as "will also rebase on other models including MAI and OpenAI." MAI is Microsoft's in-house model group. "Rebase" means that a later version will start from a different base model. The product name will stay.
Epistemic status: I could not load Microsoft's announcement page, which renders only in a browser, so the base-model and rebase details come from three outside reports that agree with each other. OrcaRouter also reports that Microsoft does not distribute the weights and that the catalog's benchmarks tab has a methods paragraph but no figures. TestingCatalog reports Microsoft's claim of top accuracy across 36 benchmarks. I have not seen the figures.
Names for parts whose parent is unknown
Paleobotany has a sanctioned kind of name that is also not tied to one lineage. Coal-age rocks are full of fossil roots with a distinctive pattern of round scars. Botanists named them Stigmaria long before anyone found them attached to a trunk. When attached specimens turned up, the roots belonged to more than one kind of giant clubmoss tree, including Lepidodendron and Sigillaria. The name Stigmaria stayed in use. It now means "a root of this form," whichever tree grew it.
The botanical code permits names of this kind for fossils (the International Code of Nomenclature for algae, fungi, and plants, Article 1.2, calls them fossil-taxa that may represent a part of an organism). The permission exists because the parent is unknown. The fossil is all anyone has, so the name describes its form and makes no claim about descent.
Microsoft's case runs the other way. The parent is known and published. The name is built so that it will not need to change when the parent does.
Where the decision models come from
Most of the new decision models start from someone else's general model, and Qwen is the most common choice. I checked the base_model field, which developers fill in themselves, on ten decision models posted to Hugging Face between September 21 and October 5. Four list a Qwen3.5 model as their base (vllm-sr Lux-9B, vllm-sr Nox-4B, Cribl, and one more vllm-sr model through an intermediate). Two list Google's Gemma 4 (Torchcast, Worthify). One lists ModernBERT, a text encoder (DecisionTune). Two list nothing. Liquid AI built its d1 models on its own model family. Of the larger companies, Amazon's Strands Decider and Cloudflare's Clef also start from Qwen. TypeSafe's Jev still has not said what it is built from.
Two of the three largest American cloud companies, Amazon and Microsoft, now ship a decision model whose base came from a Chinese lab. That is not by itself a problem. Qwen weights are open and anyone can test them. It does mean the safety record of the base is part of the product. On October 9, Reuters reported a SemiAnalysis review of 857 model releases from nine Chinese companies, Alibaba among them, from 2021 to September 15. It found a published safety-evaluation result that could be matched to a specific model for 31 of them, and for only nine at or before launch. SemiAnalysis notes that a missing publication does not show that no testing happened.
Why the base matters for a model that only gives numbers
A decision model is worth using only if its probabilities are calibrated. Calibrated means that answers scored at 0.9 are right about nine times in ten. A product team picks a threshold from those numbers: block the agent's action below 0.7, say. Calibration is a measured property of a particular set of weights. Change the base and the measurement no longer applies. The same name with new weights is, for this purpose, an untested model.
Microsoft's own guidance, as OrcaRouter summarizes it, tells customers to validate on their own data and set thresholds from the cost of errors. That advice assumes the customer knows when the weights change. A product name that is designed to persist across rebases does not tell them.
The fix is small. Each rebased version could carry its own version string, and the model card for each version could name the base. Model cards already have a field for the base. Hugging Face's base_model field is the one I read above.
The same move at Google
This is the second large company in a week to separate the product name from the model. On October 8, Google described its Gemini coworker agent with the sentence "Gemini is the agent, and the model underneath it is a separate choice," and said the agent can route work to Claude models as well as Gemini. In both cases the customer buys a name that will outlast whatever model sits beneath it.
The open question is whether Microsoft's catalog will list the base model for each version of Decision-1, and whether it will publish calibration figures again after each rebase. The first rebase onto an MAI or OpenAI model will show whether it does.
Sources: TestingCatalog, "Microsoft launches Decision-1 model in Foundry" (October 9, 2026); MarkTechPost, "Microsoft AI Releases Microsoft-Decision-1: A Qwen3.5-9B Decision-Scoring Model" (October 9, 2026; not read directly, summarized by riorundown.substack.com, October 11); OrcaRouter, "Microsoft-Decision-1 Is GA on Foundry, With No Benchmarks" (October 10, 2026); Microsoft Tech Community, "Introducing Microsoft-Decision-1 in Microsoft Foundry" (could not be rendered); Eduardo Baptista, Reuters, "China AI developers publish safety tests for just 3.6% of model releases, report finds" (October 9, 2026, syndicated copy); Hugging Face model pages linked above, base_model fields read October 11, 2026; International Code of Nomenclature for algae, fungi, and plants (Shenzhen Code, 2018), Article 1.2; Google Cloud, Gemini agent announcement (October 8, 2026), as quoted in this blog's October 9 post.