Make AI Yours: Introducing Alpha v0.2

What would AI look like if more people had a say in how it’s built, not just how they use it?

AI Potluck brings people and organizations together to build a public-interest alternative, combining contributions from across the open AI ecosystem. Alpha is where you can try that approach in practice: a working product, built together, that helps us discover what’s possible and what still needs work.

Today we’re introducing Alpha v0.2. Building on the ethical AI assistant we launched in July, this release moves us closer to AI that is yours. It focuses on three priorities: AI that works in your language, gives you control over how your data is stored and used, and lets you decide everything from its tone to what it remembers.

Your language

The language you speak should not determine whether AI is useful to you. Yet while the world has more than 7,000 spoken languages, most AI systems are trained on English and a handful of other dominant languages. Around 96% of languages are spoken by fewer than a million people each, and these remain systematically underrepresented in AI infrastructure and applications.

With v0.2, we’ve begun to address this gap. We’re expanding Alpha’s multilingual capabilities and the ways communities can contribute to evaluating them. That means looking beyond whether a model can produce an answer in a language, and asking how well that answer works for the people who speak it.

Here's what this looks like in practice:

Chat with Alpha in any language.

Alpha’s built-in model picker detects the language you’re writing in and automatically switches the model behind your chat to help you get a better response.

To inform those choices, our partner The Collective Intelligence Project (CIP) tested each available model across 122 languages using Belebele, a multilingual reading-comprehension benchmark. Those results help Alpha select a model based on how it performs in your language, not just its overall performance. Reading comprehension is one measure of language capability, not a complete picture of how well a model will work for you.

The model picker is powered by Clef, Cloudflare’s open source decision model, which pairs the Belebele scores with the Artificial Analysis Intelligence Index in order to suggest the best possible model for your language capabilities.

If your language is underserved by the available models within Alpha, an assistant pops up to help you choose your preferred model. The assistant also offers pathways for improving the datasets available in your language — either by contributing your voice via Common Voice, by contributing a dataset, or simply making a request via the Mozilla Data Collective's AI Potluck landing page.

Our partner Taraaz contributed an in-app language evaluation tool, with questions adapted from the Multilingual AI Safety Evaluation Lab. Roya Pakzad developed the lab during her Mozilla Foundation fellowship to evaluate how AI answers change across languages. In Alpha, you can use the tool to rate anonymized model responses across Italian, Portuguese and French languages, with many more to come.

Your data

Making AI yours also means understanding and having more control over what you upload, what gets stored, and what you can delete.

In Alpha v0.2, signed-in account holders can upload and store files (up to 250MB). Your files are stored on Hugging Face, in a storage bucket owned by Current AI. If you prefer, you can also connect your own Hugging Face bucket so your files are stored in your own environment.

From My files, you can view everything you’ve stored and delete any file at any time.

Alpha’s Data and privacy page explains where your data is kept and for how long. You can erase all your files or data at anytime without deleting your account.

In Account, you can download a ZIP export of your conversations and files, or export your conversations to use with Thunderbolt.

Your rules

People use AI for different purposes. One set of defaults won’t work for everyone.

Alpha v0.2 gives you more control over how Alpha responds. You can adjust its tone, answer length, suggestions, answer language, plain language, format, emojis, and how it refers to itself. You can also choose whether to show the model and host. These settings apply from your next message, in new and existing chats; Agent mode follows only the answer-language setting.

Coming soon, you’ll also be able to control what Alpha remembers from your chats to personalize future ones, including whether chats that used web search are included, and erase everything Alpha has remembered.

These settings are table stakes, and over time will evolve to work for communities and teams collaborating in a shared workspace.

Your safety

Safety takes more than a single model. In Alpha v0.2, we’re combining tools from Zentropi, Mistral and ROOST with safety policies tested against open datasets.

Zentropi’s CoPe-S1 screens chat messages before a response is generated, while Mistral’s Shieldstral screens messages in the agent service. ROOST’s Coop gives human reviewers a place to review flagged content, connecting automated checks with human oversight.⁠⁠

Each chat message goes through a general safety check and a separate check for critical risks. When the result is unclear, an additional check assesses the message against each of our 16 safety policies. If any check blocks the request, the conversation stops.

To develop those policies, we tested around 80 candidates across approximately 75 open datasets. But coverage is uneven. Tests show stronger performance in the African languages evaluated, with gaps remaining in some European and Asian languages, partly because open safety datasets are missing.

These checks reduce risk; they don’t guarantee that every response will be safe or free from bias. Better multilingual safety needs more shared data, testing and contributions from across the open AI ecosystem.

Open to your contribution

Alpha is a working model of what it looks like to build an AI product that puts people before profit, using only open source contributions.

Try it, identify its limits, and contribute to the build to help us take Alpha from a demo to our vision of a fully open source, public option for AI.

Explore the release, try it in your language, and help us understand what would make it more useful to you.