Microsoft builds its decision model on Qwen, not OpenAI
Sunday's haul mixed one real model launch with a milestone for China's driverless delivery fleets — and a robotics round that's really about drug-making capacity.
Microsoft launched Microsoft-Decision-1, a fast decision-scoring model built by post-training Alibaba's Qwen3.5-9B. Unlike an LLM that generates text, a decision model returns a calibrated probability — the kind of call an agent makes when it routes a ticket, flags an incident, or approves a refund. Microsoft says the model takes first place in its own 36-benchmark comparison covering nearly 150,000 questions, answers at 85 milliseconds at the median, and costs $0.042 per million input tokens with output free. It's available now in Foundry and on OpenRouter, and Microsoft frames it as the first in a new class of models it will soon rebase on its own MAI models and on OpenAI's.
Every headline number is Microsoft's own, and none have been independently verified — including the claim that it runs 35 times faster at median latency than GPT-6 Sol. H2O.ai has already disputed Microsoft's latency comparison, arguing that adjusted figures for its own model double from 29 milliseconds to 210 milliseconds, while Microsoft's model doesn't appear on the JevBench leaderboard at all. The interesting part isn't the leaderboard, it's the base model: Microsoft shipped a flagship release on an open-weight model from a Chinese lab, which tells you Qwen still powers products inside its rivals.
Neolix now runs what Bloomberg calls the world's largest driverless delivery fleet — 27,000 robovans across 300 cities in 15 countries. The Beijing company says that's nearly five times the combined size of Waymo's and Baidu's robotaxi operations, and it's targeting 50,000 domestic vehicles this year. The operational milestone is Shenzhen, which opened nighttime driverless delivery in March 2026: after-dark routes there have grown from 2 to 437, with more than 170 vans moving parcels overnight while 42 vehicle metrics stream into a municipal monitoring platform. One caution — the 27,000 figure is Neolix's own, and no outlet outside Bloomberg has independently verified it.
Multiply Labs raised $75 million to robotize drug manufacturing — a Series B led by Patrick Soon-Shiong's NantWorks, with AstraZeneca, Lingotto and Teradyne joining. The company's enclosed robotics clusters imitate human manufacturing steps, so the output doesn't require fresh regulatory approval; Multiply claims a 100x throughput boost and 74% lower cost per dose. CEO Fred Parietti's framing is the story: "AI is designing more therapies than the industry will ever be able to manufacture." With more than $100 million raised to date, the bet is that the AI era's bottleneck is physical capacity, not model capability.
What to watch: whether the promised MAI- and OpenAI-based versions of Decision-1 actually ship, and whether its latency claims survive independent benchmarking now that H2O.ai is contesting them.
Is a 9B model that scores decisions cheaper and faster than a frontier LLM the more meaningful release? Tell us in the comments.
