We are releasing something important with full transparency 🧵
Fikira Dataset by @VamboAI : a synthetic reasoning dataset for 10 African languages.
Version 1.0 is available now:
🔗 huggingface.co/datasets/vamb…
Here is why we are doing this and what we need from you:
Most African LLMs are Western models with a sprinkle of Swahili.
Vambo AI just released MORENA: a 1.5B model trained from scratch for 12 African languages. Not a fine-tune. Open weights. Commercial use allowed.
It beats models from Google, Meta, Alibaba and Hugging Face on African language modelling and translation while being up to 8× smaller.
Training compute: about $40k of GPUs.
Shona, Swahili, Hausa, Yoruba, Igbo, isiZulu, isiXhosa, Kinyarwanda, Setswana, Afrikaans, isiNdebele, Nigerian Pidgin.
This is the part people keep saying Africa “can’t do.”
Weights: huggingface.co/collections/v…
Announced by @thisisisheanesu
/ @VamboAI
$40,000 worth of GPUs
𝗧𝗵𝗶𝘀 𝗶𝘀 𝗠𝗢𝗥𝗘𝗡𝗔: the largest African language model ever trained from scratch, at 1.5 billion parameters. Not a fine tune, this was trained from the ground up for Africa. 🌍
Trained on 12 African languages, it beats models from Google, Meta, Alibaba and HuggingFace on African language modelling and translation while being up to 8x smaller.
𝗠𝗢𝗥𝗘𝗡𝗔 doesn't belong to us. The weights are open and free to use commercially, so researchers can build on a base that doesn't treat African languages as an afterthought.
Weights in the comments.
Thanks to @UNDP, @AIHub4SD and CINECA for the GPUs and the support. None of this happens without them.
We are proud to release 𝗠𝗢𝗥𝗘𝗡𝗔 today and contribute another foundation for building AI that understands and works with African languages.
African languages deserve world class AI.
Welcome to 𝗠𝗢𝗥𝗘𝗡𝗔. 👑🌍
But is about more than the numbers.
Most African language models begin by adapting an existing large language model. 𝗠𝗢𝗥𝗘𝗡𝗔 starts from scratch, with its tokenizer, mixture and target languages designed specifically around African languages from the outset.
On held-out African language text, 𝗠𝗢𝗥𝗘𝗡𝗔 achieves the lowest bits per byte of all 26 models measured, including an 8B African specialist and a 12B general purpose model.
𝗠𝗢𝗥𝗘𝗡𝗔 1.5B base: 1.408 bpb
𝗠𝗢𝗥𝗘𝗡𝗔 1.5B instruct: 1.441 bpb
🚨 Introducing 𝗠𝗢𝗥𝗘𝗡𝗔
Today, we are releasing 𝗠𝗢𝗥𝗘𝗡𝗔, a new multilingual AI model built for African languages.
𝗠𝗢𝗥𝗘𝗡𝗔 is a 1.5 billion parameter language model trained from scratch for 12 African languages, with its own tokenizer designed for Africa.
$40,000 worth of GPUs
𝗧𝗵𝗶𝘀 𝗶𝘀 𝗠𝗢𝗥𝗘𝗡𝗔: the largest African language model ever trained from scratch, at 1.5 billion parameters. Not a fine tune, this was trained from the ground up for Africa. 🌍
Trained on 12 African languages, it beats models from Google, Meta, Alibaba and HuggingFace on African language modelling and translation while being up to 8x smaller.
𝗠𝗢𝗥𝗘𝗡𝗔 doesn't belong to us. The weights are open and free to use commercially, so researchers can build on a base that doesn't treat African languages as an afterthought.
Weights in the comments.
Thanks to @UNDP, @AIHub4SD and CINECA for the GPUs and the support. None of this happens without them.
@VamboAI got in!🌟
The amount of hard work and sacrifice that has happened between this picture from 2023 and this @GoogleStartups milestone deserves its own book.
Here’s to getting in! And after the 5th attempt because we keep trying until it works out🔥
I am so grateful.
@VamboAI has been building localised infrastructure for Africa with a focus on African languages. We are grateful to see our work featured in a @StanfordGSB case study on AI adoption in developing economies alongside ventures from across the world 🌍
gsb.stanford.edu/faculty-res…
Here are some more innovators shaping Africa’s AI future at #AfricaAIVillage Expo – India AI Impact Summit.
From Nigeria
1. Udu Technologies – A trusted AI partner helping governments and enterprises turn their AI ambitions into reality.
2. mDoc – A digital healthcare pioneer connecting patients and providers through remote care and monitoring solutions.
3. @equalyz_ai – Democratizing AI for Africa’s diverse language communities by building inclusive AI tools tailored to local contexts.
From South Africa
1. @OptimusRobotics – Building next-generation robotics solutions powered by advanced AI technologies.
2. UmkhoAI AI – Driving African self-reliance through AI solutions that empower governments to better serve their people.
3. @VamboAI | Multilingual Intelligence – Enabling users to access powerful AI tools in their native African languages.
From Zimbabwe
1. Justtrade365 – An AI-powered platform simplifying trade by connecting businesses to market insights, compliance support, logistics, and finance solutions.
From Zambia
1. @agripredict – Using satellite data and AI to give farmers real-time advice on weather, soil, crops, and market prices — accessible even on basic mobile phones.
From healthcare to agriculture, robotics to trade, African innovators are building AI solutions designed for real-world impact across the continent.
#AIVillageInIndia#IndiaAIImpactSummit#SouthSouthCooperation#AIinAfrica#GlobalMajority#ResponsibleAI#DigitalInfrastructure
400+ downloads in 48 hours 🎉
Fikira Dataset by @VamboAI is reaching researchers and developers building African language AI.
For those just hearing about it:
→ Reasoning data for 10 African languages
→ 50,000+ examples
→ Open source, free to use
🔗huggingface.co/datasets/vamb…
We are releasing something important with full transparency 🧵
Fikira Dataset by @VamboAI : a synthetic reasoning dataset for 10 African languages.
Version 1.0 is available now:
🔗 huggingface.co/datasets/vamb…
Here is why we are doing this and what we need from you:
400 downloads in two days 🎉
The research community has spoken, I'm working on 2.0 and hope to get more languages, more examples, and human validators.
Feedback is always appreciated 🤠
We are releasing something important with full transparency 🧵
Fikira Dataset by @VamboAI : a synthetic reasoning dataset for 10 African languages.
Version 1.0 is available now:
🔗 huggingface.co/datasets/vamb…
Here is why we are doing this and what we need from you: