Lelapa AI / InkubaLM
Africa's first multilingual SLM — 0.4B parameters, 5 African languages, outperforms models 20x its size
Overview
Lelapa AI is a Johannesburg-based AI research and product lab founded in December 2022 by CEO Pelonomi Moiloa and CTO Jade Abbott. The company builds resource-efficient language AI designed to work reliably under real-world infrastructure constraints in emerging markets. InkubaLM-0.4B, released in August 2024, is Africa's first multilingual small language model trained from scratch on 2.4 billion tokens spanning isiZulu, Yoruba, Hausa, Swahili, and isiXhosa alongside English and French — with a custom vocabulary of 61,788 tokens optimised for African morphology.\n\nOn benchmarks including AfriMMLU, AfriXNLI, machine translation, and sentiment analysis, InkubaLM matches or outperforms models with far greater parameter counts — including LLaMA 3-8B and SmolLM-1.7B on several tasks — demonstrating that African-language-first tokenisation and focused training data yield outsized efficiency gains. The model is publicly available on Hugging Face under a CC BY-NC 4.0 licence and can run on consumer hardware, including laptops without GPUs. Lelapa's commercial API product, Vulavula, extends these capabilities with speech-to-text and machine translation services targeting African enterprise and contact-centre customers.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- 1,000 API calls/mo
- Transcribe & Translate
- 10,000 API calls/mo
- Transcribe & Translate
- 500,000+ API calls/mo
- Reduced per-call rates
- No rate limits
Use cases
What you can produce with Lelapa AI / InkubaLM
- Open-weight SLM (InkubaLM-0.4B, CC BY-NC 4.0) available via HuggingFace for non-commercial research and fine-tuning
- Vulavula Transcribe API — multilingual speech-to-text with code-switching support for African languages
- Vulavula Translate API — context-aware machine translation between supported African languages
- Inkuba-Mono and Inkuba-Instruct open datasets on HuggingFace for African language NLP research
- Technical paper (arXiv preprint 2408.17024) documenting training methodology, benchmark results, and reproducibility details
- Developer documentation at docs.lelapa.ai for integrating Vulavula API into applications
ASEAN Perspective
Lelapa AI / InkubaLM in Southeast Asia
InkubaLM is not directly applicable to ASEAN use cases — its five supported languages (isiZulu, Yoruba, Hausa, Swahili, isiXhosa) have no overlap with Southeast Asian languages. However, its core methodology — training from scratch with language-first tokenisation on low-resource data budgets, achieving parity with far larger models — is directly relevant to ongoing ASEAN efforts to build indigenous language models for Bahasa Indonesia, Filipino, Thai, and other regional languages. ASEAN AI researchers and policymakers building sovereign language infrastructure will find InkubaLM's efficiency results and open datasets (Inkuba-Mono, Inkuba-Instruct on HuggingFace) a useful technical reference. As a model to use today for ASEAN tasks, it offers nothing; as a research blueprint for what a well-resourced but compute-constrained regional AI lab can achieve, it is highly instructive.
InkubaLM is a landmark achievement in efficiency-first language AI — a 0.4B-parameter model that genuinely competes with models 10–20x larger on African-language benchmarks. The research is rigorous, backed by a peer-reviewed arXiv paper, compute support from Microsoft AI4Good, and prominent ecosystem validation (Brad Smith citation, Mozilla Ventures backing). The May 2025 Buzuzu-Mavi challenge, which compressed the model a further 75% to roughly 0.1B parameters while retaining performance, underscores the team's commitment to deployability on entry-level hardware — a real differentiator for emerging-market contexts.\n\nThe caveats are material for most RECATOOLS readers. InkubaLM's five supported languages are exclusively sub-Saharan African; there is no ASEAN or APAC language coverage, and no roadmap has been announced for SEA expansion. The CC BY-NC 4.0 licence bars commercial use of the base model weights. The commercial Vulavula API covers only transcription and translation (not general text generation), starts at $9.99/month with a cap of 1,000 calls, and lacks a free tier. For teams outside Africa, InkubaLM is best understood as an important proof-of-concept for resource-efficient multilingual SLMs rather than a production-ready general-purpose tool.
What people say
$9.99 a month buys 1,000 API calls on Lelapa AI's Vulavula platform — cheap enough for a side project, but there's no free tier to test before you pay. That's the commercial reality behind InkubaLM, the 0.4B-parameter model Lelapa (Johannesburg, founded 2022 by Pelonomi Moiloa and Jade Abbott) trained from scratch on isiZulu, Yoruba, Hausa, Swahili, and isiXhosa.
The research holds up. InkubaLM matches or beats models many times its size — LLaMA 3-8B and SmolLM-1.7B among them — on AfriMMLU and AfriXNLI benchmarks, runs on a laptop CPU, and is backed by a peer-reviewed arXiv paper plus a $2.5M seed round (Mozilla Ventures, Atlantica Ventures, and Google's Jeff Dean as an individual backer). The Buzuzu-Mavi challenge in 2025 shrank it a further 75% to roughly 0.1B parameters without losing much accuracy — the kind of efficiency work that actually matters for anyone deploying on cheap hardware in low-connectivity markets.
What you can't do is put InkubaLM into a commercial product as-is — the weights ship under CC BY-NC 4.0, non-commercial only. For revenue-generating use, Vulavula is the paid path, and it only does speech-to-text and translation, not general text generation. Pricing runs Dev Pass ($9.99/mo, 1,000 calls) to SMME Pass ($49/mo, 10,000 calls) to a custom Enterprise tier for 500,000+ calls. None of the five supported languages are ASEAN or APAC languages, so for RECATOOLS readers this is a reference point for low-resource-language methodology rather than a tool to actually deploy.
Summary of public user & expert reviews, compiled by RECATOOLS.
Notable facts
- InkubaLM was trained from scratch — not fine-tuned from an existing Western model — with a custom 61,788-token vocabulary built specifically for African language morphology.
- In the May 2025 Buzuzu-Mavi global challenge (490 participants from 61 countries), African developers took all three top spots and compressed InkubaLM by 75% without meaningful performance loss.
- CEO Pelonomi Moiloa holds a Master's in biomedical engineering from Tohoku University, Japan — her neural imaging research background directly informed Lelapa's approach to low-data, high-efficiency model design.
- Microsoft President Brad Smith specifically named Lelapa AI in a March 2025 speech on AI in South Africa, citing it as a model for African-led AI infrastructure development.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Lelapa AI / InkubaLM's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Lelapa AI / InkubaLM unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Lelapa AI / InkubaLM directly →
Spotted something out of date? Suggest an update →
Alternatives to Lelapa AI / InkubaLM
More in LLMs & Chat
