Most spoken-language AI works in stages. Audio comes in, a speech recognition model turns it into text, a language model reads that text and generates a reply. Three steps, at least, often across three separate models, usually requiring a live connection to whichever server is running them.
That approach breaks down fast in a classroom with no internet connection. For an offline Android app that lets a child ask a question out loud in Luganda or Acholi and get a spoken-language answer back, Sunbird AI needed something that could do the whole job on the phone itself, in one pass, with nothing sent anywhere.
No Separate Transcription Step
Sunflower v2 Gemma4-E2B is a text/speech model based on Google’s Gemma 4’s E2B model and fine-tuned on Sunbird AI’s language data, built specifically to address this challenge.. Gemma 4 is a natively audio model, meaning it can take raw audio directly as input, not just text. Sunbird AI’s fine-tuning takes advantage of that directly; 16kHz mono audio goes straight into the model, with no separate transcription step in between.
That matters more than it might sound. The audio understanding and the answer generation happen inside the same context window, in a single forward pass. Ask a question out loud, get an answer back, with no intermediate cascade of models handing work off to each other. Depending on the prompt, the same model can answer the question directly, transcribe it, translate it into another supported language, or explain what was said.
Built for the Classroom, Not the Data Center
The target environment shaped the whole approach. Sunbird AI’s Sunflower v2 Gemma4-E2B model runs entirely offline on Android. A primary school child in a Ugandan classroom can ask a science question in Luganda and hear the answer back in Luganda, with no data plan, no signal, and no server round trip. Aimed at exactly this kind of low-connectivity, real-world setting rather than a demo built for ideal conditions.
How Well It Performs
English is the strongest language, as expected for a model built on Gemma 4. Among the [20] Ugandan languages it was fine-tuned on, Luganda performs best, scoring roughly 0.51 chrF on the project’s evaluation set. Acholi is second tier, at roughly 0.40 chrF, usable in the classroom for shorter responses rather than long, complex ones.
Performance is influenced by the amount of training data available per language and the capacity of the base model, which is consistent with what Sunbird AI has found across its other models.
Open, Including the Training Process
The model weights are open and published on Hugging Face. The training pipeline, data preparation, and evaluation harness used to build it are also public, for anyone who wants to see exactly how it was built or adapt the approach for another language.
Part of a Larger Suite
Gemma4-E2B is the offline, on-device model in the Sunflower v2 suite of models released on September 24, 2026, alongside a flagship reasoning and translation model, a speech recognition model, and a text-to-speech model.