Get 40% Off GitKraken Pro
Get 50% Off Ends in
Days
Hours
Minutes
Seconds

Git Blog

Releasing the Power of Git

What Are Stealth Models?

The recent mystery around Ox Alpha last week and the week before was a fun slice of what makes social media fun. The hunt for the model provider and how people did that discovery should be studied. But, this post is more about stealth models in general. You might be wondering what the phrase “stealth model” even means and if you already know, you might still be curious about why companies release stealth models.

The Basics

Stealth models are models that are released publicly for short periods of time under pseudonyms. Traditionally, they have been released on OpenRouter and a few other places but the recent Ox Alpha made it to OpenCode as well. As a quick aside, OpenCode being added as a potential stealth model host is significant and highlights how much growth the OpenCode team has been seeing lately. However, OpenRouter is the main place you will see stealth models released still.

The History of Stealth Models on OpenRouter

As of today, I count 19 stealth models that have been soft-released on OpenRouter. Yes, that link shows 21, but 2 of them are not true stealth models. “Alpha” was just the best way to search for previous stealth models because that has been the naming convention thus far. Many have since been revealed and the list is pretty impressive. In order of release:

  • Sonoma Sky and Dusk Alpha: Grok 4 Fast
  • Andromeda Alpha: NVIDIA Nemotron Nano 2 VL
  • Polaris Alpha: GPT 5.1
  • Sherlock Think and Dash Alpha: Grok 4.1 Fast
  • Bert-Nebulon Alpha: Mistral Large 3
  • Pony Alpha: GLM 5
  • Healer Alpha: MiMo v2 Omni
  • Hunter Alpha: MiMo v2 Pro
  • Elephant Alpha: Ling 2.6 Flash
  • Ox Alpha: GLM 5.3 Flash

Why Release Stealth Models?

The reasons for releasing a model as a stealth model will differ from company to company. But, in general, it gives the company a chance to have the model evaluated by a large number of people regardless of subscriptions. That means something like Polaris Alpha can get into the hands of many more people than just the subscribers to OpenAI. The amount of feedback and telemetry they can gain from this more than makes up for the cost of offering it for free. It’s also free marketing without the baggage attached to people’s perceptions of a specific company.

Not Just Marketing

If you look at the list above, you might notice that it was all 19 models that are listed as previous Alphas on OpenRouter. Beyond just marketing, companies can evaluate if a model is even worth releasing. 8 models in the full list are not publicly disclosed as what model it really was. If a company finds that the performance or cost of running it is too high, they can just quietly disable the model without having an impact on the company itself.

But Why Ox Alpha Specifically?

The speculation around Ox Alpha before it was revealed was so much fun. A lot of social media accounts jumped on the hunt. Several providers haven’t released a new model in a while and were part of the suspects.

Some people thought it would be a new MiMo model since they have done stealth models in the past. Others thought maybe a new Minimax model but that seemed less likely since they haven’t done the stealth model thing before. OpenAI just released a model so it seemed unlikely to be them.

Several people pointed out that the tokenizer was the same as GLM models but it seemed odd that z.ai would release a new GLM model since GLM 5.3 had just released a week prior. That didn’t prevent the polymarket odds from being heavily in favor of a new GLM model, though. Ox Alpha was also a multi-modal model which allowed images, video and text while GLM 5.3 is just text.

That is where the “why” starts to make sense. It was z.ai’s first attempt at multi-modal input and they could have gotten it wrong. They needed broad testing of all its capabilities and they sure got it.

What’s Next?

With Ox Alpha revealed, it remains to be seen what the next stealth model release will be. If we consider the usual suspects, we might see a new MiMo release soon. If they release 2 stealth models at the same time, then the odds are heavily in their favor. But, Mistral is also somewhat silent lately. Who knows? Maybe we’ll see something out of a new player. Either way, I look forward to seeing the sleuthing on social media.

Like this post? Share it!

Read More Articles

Visual Studio Code is required to install GitLens.

Don’t have Visual Studio Code? Get it now.

Team Collaboration Services

Secure cloud-backed services that span across all products in the DevEx platform to keep your workflows connected across projects, repos, and team members
Launchpad – All your PRs, issues, & tasks in one spot to kick off a focused, unblocked day. Code Suggest – Real code suggestions anywhere in your project, as simple as in Google Docs. Cloud Patches – Speed up PR reviews by enabling early collaboration on work-in-progress. Workspaces – Group & sync repos to simplify multi-repo actions, & get new devs coding faster. DORA Insights – Data-driven code insights to track & improve development velocity. Security & Admin – Easily set up SSO, manage access, & streamline IdP integrations.
winget install gitkraken.cli