<?xml version="1.0"?>
<!--<!DOCTYPE nitf SYSTEM "nitf-3-4.dtd">-->
<nitf>
  <head>
    <title id="Title">#Title</title>
    <docdata management-doc-idref="">
      <date.issue id="CreationDate" norm="" />
      <du-key id="rev-ver" generation="1" version="Default" />
      <du-key id="Parent-Version" version="" />
      <identified-content>
        <classifier id="newspro-nitf" value="r2" />
        <classifier id="Newspro-App" value="Epaper" />
        <classifier id="Content-Type" value="Story" />
        <classifier id="storyID" value="423469" />
        <classifier id="CmsConID" value="" />
        <classifier id="Desk" value="" />
        <classifier id="Source" value="" />
        <classifier id="Edition" value="" />
        <classifier id="Category" value="-1" />
        <classifier id="UserName" value="" />
        <classifier id="PublicationDate" value="13/10/2026" />
        <classifier id="PublicationName" value="DailyStar" />
        <classifier id="IsPublished" value="Y" />
        <classifier id="IsPlaced" value="Y" />
        <classifier id="IsCompleated" value="N" />
        <classifier id="IsProofed" value="N" />
        <classifier id="User" value="" />
        <classifier id="Headline-Count" value="" />
        <classifier id="Slug-Count" value="0" />
        <classifier id="Photo-Count" value="0" />
        <classifier id="Caption-Count" value="0" />
        <classifier id="Word-Count" value="0" />
        <classifier id="Character-Count" value="0" />
        <classifier id="Location" value="" />
        <classifier id="TemplateType" value="1" />
        <classifier id="StoryType" value="Story" />
        <classifier id="Author" value="" />
        <classifier id="UOM" value="mm" />
        <classifier id="kicker" value="" />
        <classifier id="ByLine" value="" />
        <classifier id="DateLine" value="" />
        <classifier id="box-geometry" value="660,1095,950,1490" />
         <classifier id="Layer" value="Layer 1"/>
          <classifier id="numcol" value="2"/>
         <classifier id="ArticleStyle" value=""/>
       <classifier id="Epaper-Build" value="7.1.5.2"/>
  <classifier id="ProcessingDateTime" value="Mon Oct 12 2026 19:19:13 GMT+0600"/>
      </identified-content>

      <urgency id="home-page" ed-urg="0" />
      <urgency id="priority" ed-urg="0" />
      <doc-scope id="scope" value="0" />
    </docdata>
    <pubdata type="print" name="DailyStar" date.publication="20261013T000000+5.30" edition.name="Dhaka Edition" edition.area="MAI" position.section="DST13102610MAI-NEXTSTEP" position.sequence="10" ex-ref="DST13102610MAI-NEXTSTEP.indd" />
  </head>
  <body boxBorderWeightColor="" boxBorderWeight="">
<body.head>
      <hedline>
    	<hl1 id="Headline1" ul="0" ol="0" ulColor=""  ulWeight=""  olColor=""  olWeight="" textFrameColor="" orgstyle="HEAD new 25 HEAVY" class="1" MainHead="true" style="Headline1">
		<lang class="3" colour="#000000" orgstyle="HEAD new 25 HEAVY" style="Headline1"  font="Antonia H1" fontStyle="Regular" size="22">Microsoft’s new AI model scores decisions instead of writing text</lang>
	</hl1>

       </hedline>
</body.head>
    <body.content id="Bodytext" CaptionAsBody="0">
     <block>
	<media id="1" media-type="image">
		<media-reference id="tn" source-credit="" data-location="1" source="MicrosoftAImo_10_MAI_tn.jpg"  Units="pixels" width="50" height="50"></media-reference>
	</media>
</block>

     <p style=".Bodylaser" ul="0" ol="0"  orgstyle="NAME new 2025">
	<lang class="3" style=".Bodylaser" colour="#000000" orgstyle="NAME new 2025" font="Antonia H1" fontStyle="Medium" size="7">IMAGE:</lang>
<lang  class="3" style=".Bodylaser" colour="#000000" orgstyle="NAME new 2025" font="Antonia H1" fontStyle="Heavy" size="7"> REUTERS</lang>
</p>
<p style=".Bodylaser" ul="0" ol="0"  orgstyle="BY NAME_2025">
	<lang class="3" style=".Bodylaser" colour="#000000" orgstyle="BY NAME_2025" font="Impact" fontStyle="Regular" size="9">▌</lang>
<lang  class="3" style=".Bodylaser" colour="#000000" orgstyle="BY NAME_2025" font="Antonia H1" fontStyle="Bold" size="9">	NEXT STEP DESK
</lang>
</p>
<p style=".Bodylaser" ul="0" ol="0"  orgstyle="INDENTLESS BODY 2025">
	<lang class="3" style=".Bodylaser" colour="#000000" orgstyle="INDENTLESS BODY 2025" font="Antonia H1" fontStyle="Medium" size="9">Microsoft has released a decision-scoring model that returns probability scores for a fixed set of answer options rather than generating text. The model, named Microsoft-Decision-1, is available through a hosted API in Microsoft Foundry, and its weights are not distributed.
</lang>
</p>
<p style=".Bodylaser" ul="0" ol="0"  orgstyle="BODY 2025">
	<lang class="3" style=".Bodylaser" colour="#000000" orgstyle="BODY 2025" font="Antonia H1" fontStyle="Medium" size="9">When given a context, a question, and a defined list of possible answers, the model produces a calibrated probability for each option. It handles yes/no questions, multiple choice, rating scales, classification, and rubric-based questions, and supports an explicit abstention option such as “cannot tell” when the evidence is insufficient. It accepts up to 32,768 input tokens in a single pass and returns results in JSON.
</lang>
</p>
<p style=".Bodylaser" ul="0" ol="0"  orgstyle="BODY 2025">
	<lang class="3" style=".Bodylaser" colour="#000000" orgstyle="BODY 2025" font="Antonia H1" fontStyle="Medium" size="9">The model is built on Alibaba’s open-weight Qwen3.5-9B and post-trained by Microsoft, which said training data came from publicly available datasets and synthetic material, with no customer data used. Microsoft added that it plans to rebase the model on others, including MAI and OpenAI systems.
</lang>
</p>
<p style=".Bodylaser" ul="0" ol="0"  orgstyle="BODY 2025">
	<lang class="3" style=".Bodylaser" colour="#000000" orgstyle="BODY 2025" font="Antonia H1" fontStyle="Medium" size="9">The company is explicit about what the model is not for. It does not generate text, answer open-ended questions, hold conversations, translate, or summarise. It is text-only and produces no explanations. Microsoft also states it should not be the sole automated decision-maker in consequential matters involving credit, employment, housing, etc., and must not be used for surveillance, profiling, </lang>
</p>

    </body.content>
  </body>
</nitf>