Microsoft

Microsoft's own model index lists seven cards — MAI-Transcribe-2, MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6 and MAI-Voice-2 under "Foundational model", MAI-Cyber-1-Flash under the same label, and Microsoft Frontier Tuning under "Custom"source, accessed 2026-09-06. This catalog's microsoft/ namespace lists two rows, and neither is one of them. That is not a lag in the feed. On 2 June 2026 Mustafa Suleyman wrote "Today we are announcing a family of seven new models developed in-house at Microsoft AI. Beyond these models, we're building a superintelligence lab – a system and an approach we believe will define the next phase of AI." — Mustafa Suleyman, under the dateline "June 2, 2026" and the standfirst "Updated as of June 8, 2026."source, accessed 2026-09-06, and named the shelf this catalog reads from: "Alongside distribution on Foundry and optimization for our 1P products, our models are also going to be widely available for developers on OpenRouter, as well as Fireworks and Baseten. For the first time developers will be able to tune the weights of the model themselves."source, accessed 2026-09-06. Three months and four days later, The OpenRouter models API returns no `mai` or `microsoft-ai` namespace rows. Across all 431 rows the only ids or names matching `mai`, `microsoft`, `phi`, `maia` or `wizard` are microsoft/phi-4, microsoft/wizardlm-2-8x22b and an unrelated Venice fine-tunesource, accessed 2026-09-06. The line itself has not stood still in the interval — the June announcement names MAI-Image-2.5, MAI-Code-1-Flash and "MAI Transcribe-1.5"; the model index today names MAI-Image-2.6, MAI-Code-1.1-Flash and MAI-Transcribe-2source, accessed 2026-09-06 — so what has not arrived is the distribution, not the models.

Microsoft is on this router, though, at scale, under a different name. "Azure", one of the 106 entries the OpenRouter providers API returns, carrying `privacy_policy_url` https://www.microsoft.com/en-us/privacy/privacystatement and `status_page_url` https://status.azure.com/source, accessed 2026-09-06. Sweep the endpoints and the shape is unambiguous: 34 rows list Azure as an endpoint provider and every one of them is in the `openai/` namespace, from openai/gpt-3.5-turbo-0613 to openai/gpt-6-astra-pro — measured by calling /api/v1/models/<id>/endpoints once for each of the 123 rows this list carries in the `openai/`, `microsoft/`, `meta-llama/` and `mistralai/` namespaces. Neither `microsoft/` row is among the 34source, accessed 2026-09-06. Its own two rows are served by strangers — exactly one endpoint, "DeepInfra | microsoft/phi-4", quantization bf16source, accessed 2026-09-06 and exactly one endpoint, "Novita | microsoft/wizardlm-2-8x22b", quantization bf16source, accessed 2026-09-06. In the microsoft/ namespace Microsoft is the vendor of record and nothing else; the compute belongs to DeepInfra and Novita, and the compute Microsoft does sell here runs somebody else's weights.

The older of the two rows is odder than being old. Follow its declared hugging_face_id and you get HTTP 401 with the body {"error":"Invalid username or password."}, for the repository id the OpenRouter row declares as its `hugging_face_id`. The repository page at https://huggingface.co/microsoft/WizardLM-2-8x22B answers 401 as well; so does microsoft/WizardLM-2-7Bsource, accessed 2026-09-06. Ask Hugging Face for every Microsoft repository whose name matches "wizard" and you get an empty JSON array — Hugging Face lists no model repository at all under the `microsoft` author whose name matches "wizard"source, accessed 2026-09-06. This is not a common kind of rot, either. Check the same pointer on every row that declares one and this is microsoft/WizardLM-2-8x22B is the only dead upstream. 179 of the 431 rows this list returns declare a `hugging_face_id`; 178 of those ids answer HTTP 200 from https://huggingface.co/api/models/<id> and microsoft/WizardLM-2-8x22B alone does not — swept one id at a time on 2026-09-06source, accessed 2026-09-06. The weights a buyer is actually renting exist in public as alpindale/WizardLM-2-8x22B — author `alpindale`, createdAt 2024-04-16T02:36:59.000Z, license apache-2.0, 140,620,634,112 safetensors parameters in 59 shards, 416 likessource, accessed 2026-09-06 — a copy, on a stranger's account, under a licence Microsoft chose but on a repository Microsoft does not control. It even carries its own receipt: "url: https://huggingface.co/microsoft/WizardLM-2-8x22B", "branch: main", "download date: 2024-04-15 16:48:15" — the first three lines of the mirror's huggingface-metadata.txt, above a sha256 for the repository and one for each of the 59 safetensors shardssource, accessed 2026-09-06.

That timestamp is the interesting artefact, because it is the takedown's negative. 404 Media reported at the time that "Then it deleted the model from the internet a few hours later because, as The Information reported, it “accidentally missed” required “toxicity testing” before it was released." — 404 Media, 23 April 2024, reporting The Information; Microsoft "declined to comment"source, accessed 2026-09-06, and that "However, as first spotted by Memetica, in the short hours before it was taken down, several people downloaded the model and reuploaded it to Github and Hugging Face" — 404 Mediasource, accessed 2026-09-06. The mirror's metadata file pins that "short hours" phrase to a minute and signs it: a shard-by-shard checksum of what Microsoft published before it decided it should not have. Twenty-eight months on, the announcement page carries no notice of any of it: dated "Apr 15, 2024", still reading "We introduce and opensource WizardLM-2" and "New family includes three cutting-edge models: WizardLM-2 8x22B, WizardLM-2 70B, and WizardLM-2 7B", and still stating "The License of WizardLM-2 8x22B and WizardLM-2 7B is Apache2.0"source, accessed 2026-09-06 — offering, in the present tense, a licence for two repositories that no longer answer. OpenRouter's own catalog copy, not Microsoft's, still tells buyers "WizardLM-2 8x22B is Microsoft AI's most advanced Wizard model. It demonstrates highly competitive performance compared to leading proprietary models, and it consistently outperforms all existing state-of-the-art opensource models."source, accessed 2026-09-06, which was arguable in April 2024 and is now a sentence about a model its author has withdrawn.

Eight months after that, the next Microsoft weights to reach this catalog arrived with the omission answered in the open. The phi-4 card does not claim safety work in the abstract; it names the team and the scenarios: "For qualitative safety evaluation, we collaborated with the independent AI Red Team (AIRT) at Microsoft to assess safety risks posed by `phi-4` in both average and adversarial user scenarios."source, accessed 2026-09-06. Read the two rows in order and the second one's model card is visibly the first one's post-mortem — which is the only part of this pair a buyer can check, since the process failure itself was reported by a trade publication and never described by Microsoft. The rest of that card is unusually forthcoming about cost, too: "1920 H100-80G" GPUs, a "21 days" training time and "9.8T tokens" of training data, over the dates "October 2024 – November 2024", for a "14B parameters, dense decoder-only Transformer model"source, accessed 2026-09-06, released mit, the value `cardData.license` carries on the repositorysource, accessed 2026-09-06 at 14,659,507,200 safetensors parameters, all BF16source, accessed 2026-09-06.

Two small things are worth having before you spend on either row. phi-4 is not abandonware despite the date on it: nine commits on main between 2024-12-11T11:47:29Z and 2026-07-14T14:22:25Z, the last of them "Restore <|endoftext|> (100257) as a stop token in generation_config (#67)" by `gugarosa`, who also made the firstsource, accessed 2026-09-06 — a maintained artefact, not a parked one. And its age depends entirely on which record you believe, because it has three birthdays: repository initial commit 2024-12-11T11:47:29Z; the model card's stated "Release date" December 12, 2024; the OpenRouter row's `created` 1736489872, which is 2025-01-10T06:17:52Zsource, accessed 2026-09-06. A month separates the day the weights went up from the day this catalog began counting, and the board's "newest model" column reads the last of the three.