Languages
What Rezolyn speaks today, and the roadmap.
The architecture is language-agnostic — adding a language is a model + config change, not a rewrite. Coverage rolls out in phases, deepest-gap-first.
Phase 1 — available now
| Code | Language | Speakers |
|---|---|---|
yo | Yoruba | 45M |
ha | Hausa | 70M |
ig | Igbo | 44M |
sw | Swahili | 200M |
en | English | passthrough |
fr | French | passthrough |
pt | Portuguese | passthrough |
Phase 2 — Gulf & wider Africa
Amharic, Zulu, Twi, Wolof, Nigerian Pidgin, and Arabic (including Gulf dialect). Enabled per-language behind feature flags as each model is validated.
Phase 3 — South Asia & CIS
Russian, Hindi, Bengali, Tagalog, Malay.
How detection works
Every message is language-detected automatically (you don't have to tell us the language). A model classifies the input; for very short or code-switched messages a lightweight detector backs it up. If detection is uncertain, the pipeline still answers — it just leans on the LLM, which is multilingual.
See AI & model training for how per-language quality improves over time.