Microsoft's New Decision-1 Model Is Built on Alibaba's Qwen as EU Agencies Test Chinese Open-Weight AI in Isolation

Since TypeSafe AI shipped Jev in early access on Sept. 15, the "decision model" category has ballooned to more than 100 entries, according to The Register. The latest comes from Microsoft, and it runs on a Chinese foundation.

What Microsoft shipped

On Friday, Oct. 9, Microsoft unveiled Microsoft-Decision-1. Achint Srivastava, a VP of software engineering in Microsoft's Office of the CTO, announced it in a company blog post. Satya Nadella posted about it the same day.

The model does not write text. It reads an input and returns a calibrated probability for each option in a fixed set. Microsoft says it is built for routing, classification, prioritization, verification and workflow control.

Microsoft says it post-trained Alibaba's open-weight Qwen3.5-9B for single-pass decision scoring. It also says it "will soon rebase it on other models, including Microsoft AI (MAI) and OpenAI." No date has been given.

The model is listed in Microsoft Foundry and on OpenRouter, where it runs on the Decisions API rather than the chat endpoint. Pricing is $0.042 per million input tokens, with free output. OpenAI's Luna decisions endpoint charges $0.10 per million input tokens, according to the model listing.

The numbers, and who produced them

Microsoft says Decision-1 scored 83.5% average accuracy across 36 benchmarks and 147,137 questions that were kept blind from training. Its median latency was 85 milliseconds, with the 95th percentile at 125 milliseconds. GPT-6 Sol took 3.01 seconds.

Microsoft claims it is 35 times faster than GPT-6 Sol and 2.5 times faster than the runner-up, H2O-Lightning-4B v1.1. It also says it is 2.8 times faster than Jev and 4.5 times faster than Quyet-1.0-Large. On confidence scoring, Microsoft's own figures put it second to Quyet-1.0-Large at 92.2%.

These are Microsoft's measurements of its own model through Foundry. Competitor latencies come from the JevBench adjusted median, not raw timings.

H2O.ai disputes that method. Its model card says the JevBench adjustment doubles measured time and adds 0.15 seconds, and that its own median is 29 milliseconds, not 210. Decision-1 does not appear on the JevBench board. On that board's official composite, H2O-Lightning-4B ranks first at 72.5.

On safety, Microsoft ran 5,250 requests across 11 benchmarks covering harmful content, jailbreaks and prompt injection. It reports correct refusals with high retained utility but has not published a score.

On robustness, Microsoft perturbed each request eight ways, including paraphrasing and option shuffling. The model flipped its decision on 1.3% of perturbations on average, and on none when options were paraphrased, reversed or shuffled.

The Qwen question

Microsoft has not said why it started with Qwen or why it plans to switch. The Register noted that Microsoft gave no reason for the planned rebase.

The choice of base is not unusual in this category. Cloudflare's Clef models are also built on Qwen, as is H2O's model, which uses Qwen3.5-4B. Microsoft's defense is the one in its own post: it is selling the model inside "a secure, trusted environment" and plans to move to American-built foundations. Nothing in the material from Microsoft, or in anything reported about Decision-1, alleges a security flaw tied to the Qwen base.

Still, it is Microsoft's own post-training sitting on top of Alibaba's weights. Microsoft's roughly $13 billion investment in OpenAI is well documented, and the OpenAI-based version has not shipped yet.

What EU security agencies are doing with Chinese weights

A separate confirmation came last week from Brussels. ENISA executive director Juhan Lepassaar told Politico that the EU cybersecurity agency is running air-gapped tests on "a couple of Chinese open-weight models" on secure internal infrastructure. He did not name the models or the methodology.

European Commission spokesperson Maciej Berestecki told Politico that the Joint Research Centre is running a similar evaluation. It is a wider study that also benchmarks leading American and European systems, using the same air-gapped approach.

Tech Times reports that ENISA has roughly 109 staff and an annual budget of about €17 million, and says its findings carry weight in setting technical benchmarks under the EU AI Act. The outlet also lays out why the agencies avoided APIs. China's 2017 National Intelligence Law compels Chinese entities to cooperate with state intelligence work. Queries sent to a Chinese company's servers could therefore fall under it. Downloading weights and running them on hardware that never touches the internet removes that exposure. That is Tech Times' analysis, not a statement from ENISA or the Commission.

The open-weight licensing is what makes the approach possible. DeepSeek's flagship models ship under the MIT License.

What is still open

Neither agency has published findings. Neither has said which models it tested or when results will arrive. Tech Times says the results could shape how the AI Act is enforced against Chinese providers and which systems European critical infrastructure may use. That is a prediction about outcomes, not a decision.

The Microsoft model and the EU tests are separate stories, and no source ties them together. The open question is whether Microsoft's promised rebase on MAI and OpenAI models arrives, and what Decision-1's results look like when someone other than Microsoft runs the benchmarks. For now, the headline speed and accuracy figures are Microsoft's own, and H2O is already contesting the latency comparison.