Microsoft Decision-1 is Qwen: Microsoft ships 21 Qwen models, 13 Phi

Short answer: Yes, Microsoft-Decision-1 is Alibaba's Qwen3.5-9B with Microsoft post-training on top, and it is not a one-off. We pulled every model in Microsoft's official Hugging Face org: 21 declare a Qwen base and only 13 declare Microsoft's own Phi, and the Qwen-based ones pull 7.1x the downloads. The company that wrote the biggest check in Western AI builds its small models on a Chinese lab's weights, and it is right to.

TL;DR: Microsoft's Qwen-based models drew 337,637 downloads in the last 30 days against 47,360 for its Phi-based ones. Nvidia is further gone: 87 of its official models sit on Qwen and pull 9.2 million downloads a month, 55x what its Llama-based models pull. Alibaba's Qwen org alone did 245.7 million downloads in 30 days, 15x Meta's Llama org, so if you build on small models, the default base is already Chinese and the people telling you to be scared of it are shipping it.

Key numbers

What	Number	Source, checked Oct 26 2026
Microsoft models on a Qwen base	21 (337,637 downloads, 30 days)	Hugging Face API, microsoft org
Microsoft models on its own Phi base	13 (47,360 downloads, 30 days)	Hugging Face API, microsoft org
Nvidia models on a Qwen base	87 (9,244,214 downloads)	Hugging Face API, nvidia org
Nvidia models on a Llama base	50 (168,471 downloads)	Hugging Face API, nvidia org
Qwen org downloads, 30 days	245,729,309	Hugging Face API, Qwen org
Meta Llama org downloads, 30 days	16,011,269	Hugging Face API, meta-llama org
Qwen3.5-9B downloads, 30 days	6,085,485	Hugging Face, Qwen3.5-9B
Qwen3.5-9B price on OpenRouter	$0.17 in / $0.25 out per million tokens	OpenRouter models API
Decision-1 reported input price	$0.042 per million tokens	Shattered, Oct 25 2026

What Microsoft actually shipped

Microsoft-Decision-1 does not chat. It reads your content plus a fixed list of answers and returns a probability for each one in a single pass, for routing, classification, moderation and pass/fail checks inside agent pipelines.

Per Microsoft's launch post, as quoted by Shattered: "we post trained Qwen3.5-9B for fast, single-pass decision scoring and will soon rebase it on other models, including Microsoft AI (MAI) and OpenAI." Hacker News picked it up this morning.

"Will soon rebase" is doing a lot of work in that sentence. Our data says rebasing off Qwen is not something Microsoft has been doing. It has been doing the opposite.

21 Qwen, 13 Phi: Microsoft's own shelf says it

We pulled all 551 models in Microsoft's official Hugging Face org and read the base model each one declares. 503 are originals (things like mdeberta and the Phi base weights themselves). Of the 48 that are built on someone else's base, 21 name Qwen, 13 name Phi, 3 name Llama and 3 name Mistral.

The Qwen list is not research leftovers. It includes the Fara1.5 computer-use agents (4B, 9B and 27B, all on Qwen3.5), the NextCoder coding models (7B, 14B, 32B on Qwen2.5-Coder), the GUI-Actor family, the FrogBoss and FrogMini models on Qwen3, AesCode on Qwen3-VL, and UserRepGen-4B, posted October 14. The biggest is VibeVoice-Realtime-0.5B, a speech model on Qwen2.5-0.5B with 314,328 downloads this month.

Plain version: Microsoft owns Phi, a small-model line it has marketed for two years, and its own teams keep picking the Chinese base instead. That is a vote, and it is counted in commits, not press releases.

Nvidia is even further in

Microsoft is the polite case. Nvidia's official org has 1,018 models; 87 declare a Qwen base and they pulled 9,244,214 downloads in 30 days. Its 50 Llama-based models pulled 168,471. That is 55x.

Nvidia's single most downloaded model right now is nvidia/Qwen3.6-35B-A3B-NVFP4, a 4-bit Qwen build tuned for its own Blackwell chips, at 4,141,067 downloads. The company selling the shovels has decided which mine to point customers at.

Amazon has 12 Qwen-based models in its org. Apple has 6. Google, Meta and OpenAI have zero, which is exactly who you would expect: the three labs with their own open-weight families to defend.

Follow the money: who wins, who loses

Alibaba wins the default. The Qwen org did 245.7 million downloads in 30 days across its top models. Google did 108.7 million, OpenAI 52.5 million, Microsoft 35.3 million, Meta's Llama org 16.0 million. Qwen is 15x Llama. When the base is free, the lab whose weights everyone fine-tunes sets the formats, the tokenizers and the habits.

Microsoft wins the margin. Qwen3.5-9B lists at $0.17 per million input tokens on OpenRouter. Shattered reports Decision-1 at $0.042 per million input with free output. Microsoft took a free base, added a narrow skill and put it in Foundry, which is how every cheap AI product should be built.

Phi loses, and so does the scare story. Washington politicians keep calling Chinese weights a security risk. Meanwhile the two biggest US AI suppliers ship them under their own logos. Either the risk is overblown, or the people warning about it do not believe it. Our bet is the first one: weights are math you can inspect, and the panic mostly protects whoever sells the closed alternative.

How to check any model's real base in 30 seconds

You do not need to trust a launch post. Every Hugging Face model card can declare a base model, and the Hub turns it into a tag you can query.

1. Open https://huggingface.co/api/models?author=microsoft&full=true (swap in any org).
2. Look for tags that start with base_model:. The last part is the parent, for example Qwen/Qwen3.5-9B.
3. Count by family. That is exactly how we built the tables above; the full per-model list is in the data download.

One caveat we keep in view: the tag is self-declared, so a model that hides its parent counts as an original here. That only makes our Qwen numbers a floor.

Will Microsoft really rebase?

Microsoft says MAI and OpenAI bases are coming. Nobody has proven it yet, but the pattern says the rebase happens only if MAI ships a 9B that beats Qwen3.5-9B per dollar. Qwen3.5-9B was posted on February 27 2026 and has 6,085,485 downloads in the last month alone; whatever replaces it starts from zero.

And Qwen keeps moving. Nvidia already ships Qwen3.6 and Qwen3.8 builds. A rebase to a slower-moving base would be a step back that Microsoft's own customers would pay for in accuracy.

We covered the wider shift in Qwen ate open source: 58% of new fine-tunes and why IBM now calls open models good enough in IBM says open AI is 90% as good. The price side is in the LLM price war. More on the LLM hub and in AI industry.

What this means for you

If Microsoft and Nvidia pick Qwen for classifiers, routers and agents, you can stop paying frontier prices for those jobs. Take your yes/no, route-this-ticket and does-this-pass calls off the big model and onto a small open one; most teams find that is the bulk of their token bill.

Our AI Cost Audit shows which of your calls can move to a model that costs a fraction as much, using your own logs. If it does not do what it promises, we keep fixing it at no extra cost until it does.

Sources

• Hugging Face API, microsoft org (551 models, declared base_model tags), Oct 26 2026
• Hugging Face API, nvidia org (1,018 models), Oct 26 2026
• Hugging Face API, Qwen, google, openai, meta-llama, deepseek-ai, zai-org, mistralai, moonshotai orgs, Oct 26 2026
• Hugging Face, Qwen/Qwen3.5-9B, Oct 26 2026
• OpenRouter models API, Oct 26 2026
• Shattered (Tom Vance), Microsoft Decision-1: 9B Model Claims 35x Speed, Oct 25 2026
• Hacker News discussion, Oct 26 2026