Africa

AI models, vendors, and hardware buying realities from Africa. Curated by myaihardware.com.

Models

Vendors

Languages

Why Africa matters for AI builders

Africa matters for AI builders because the continent's pervasive mobile-first infrastructure means edge inference hardware — like low-power TPUs or Snapdragon SoCs — often reaches users before cloud GPUs do. The dominant model families are narrow, on-device architectures like lightweight transformer variants and diffusion models optimized for local execution, not large-scale foundation models. An underrated angle: Africa's fragmented energy grids force builders to prioritize hardware endurance and power efficiency, yielding deployment patterns that scale more resiliently than grid-reliant regions.

Market context

The development of African LLMs is constrained by severe data scarcity for low-resource languages like Wolof, isiXhosa, and Amharic, which lack the digital corpora needed for effective training. The Masakhane research community has been central to addressing this gap by coordinating grassroots efforts to collect, annotate, and translate datasets, though progress remains uneven. Sovereign AI initiatives in Rwanda, South Africa, and Kenya are now pushing for locally hosted models and regulatory frameworks to reduce dependency on foreign infrastructure, but most are still in early pilot stages. Lelapa AI, based in South Africa, has emerged as the commercial leader by releasing the first production-ready multilingual model for African languages, though its reach is limited by funding and compute constraints. The Deep Learning Indaba conference continues to serve as the key annual gathering for researchers and practitioners to share findings on these challenges, yet the gap between ambition and available resources remains wide.

Founder note · We cover what is actually buyable across African markets - accounting for customs, power-grid realities, and grey-market pricing that US/EU sites never write about.

Models (33 of 33)

Filter by vendor, license, or parameter range.

LicenseVendorParameter range

InkubaLM-0.4B

Lelapa AI · South Africa

Open

Params: 400M
Context: 2,048
Min VRAM: 1 GB
GGUF: No

Efficient small model for low-resource South African languages.

Vulavula

Lelapa AI · South Africa

Open

Params: 100M
Context: 512
Min VRAM: 0.5 GB
GGUF: No

Specialized for Xitsonga language tasks.

Afro-XLM-R Large

Davlan · Nigeria

Open

Params: 560M
Context: 512
Min VRAM: 4 GB
GGUF: No

Strong multilingual performance across many African languages.

AfriBERTa Large

Kelechi Ogueji (Castorini lab, Nigerian-led) · Nigeria

Open

Params: 110M
Context: 512
Min VRAM: 2 GB
GGUF: No

Efficient BERT-based model for multiple African languages.

AfroLM Active Learning

Bonadossou · Benin

Open

Params: 110M
Context: 512
Min VRAM: 2 GB
GGUF: No

Active learning approach reduces annotation cost for low-resource languages.

Afri-MT5 Base

Masakhane · South Africa

Open

Params: 300M
Context: 512
Min VRAM: 4 GB
GGUF: No

Multilingual T5 model fine-tuned for African language translation.

AfriQA AfroXLM-R SQuAD-v2

Masakhane · South Africa

Open

Params: 300M
Context: 512
Min VRAM: 4 GB
GGUF: No

Fusion-in-Decoder architecture for open-domain QA in African languages.

AfroXLM-R MasakhaNER 1.0 + 2.0

Masakhane · South Africa

Open

Params: 110M
Context: 512
Min VRAM: 2 GB
GGUF: No

Comprehensive NER model for 10 African languages.

AfroLlama V1

Pulse AI · South Africa

Open

Params: 7B
Context: 4,096
Min VRAM: 14 GB
GGUF: Yes

Large 7B model fine-tuned for African language generation.

Amharic RoBERTa

EthioNLP · Ethiopia

Open

Params: 110M
Context: 512
Min VRAM: 2 GB
GGUF: No

Specialized RoBERTa model for Amharic NLP tasks.

Ethio-Llama

iocuydi · Ethiopia

Open

Params: 7B
Context: 4,096
Min VRAM: 14 GB
GGUF: Yes

Large Amharic-English bilingual model for generation.

Afro-XLM-R Base

Davlan · Nigeria

Open

Params: 280M
Context: 512
Min VRAM: 2 GB
GGUF: No

Efficient base model for multilingual African NLP.

mBERT Swahili (Davlan)

Davlan · Nigeria

Open

Params: 110M
Context: 512
Min VRAM: 2 GB
GGUF: No

Specialized BERT model for Swahili NLP tasks.

JengaAI Kenya

JengaAI Kenya · Kenya

Restricted

Params: 100M
Context: 512
Min VRAM: 2 GB
GGUF: No

Focused on Swahili-English translation for Kenyan context.

BERT Hausa

Davlan · Nigeria

Open

Params: 110M
Context: 512
Min VRAM: 2 GB
GGUF: No

Fine-tuned for Hausa NER and text classification.

BERT Yoruba

Davlan · Nigeria

Open

Params: 110M
Context: 512
Min VRAM: 2 GB
GGUF: No

Fine-tuned for Yoruba NER and text classification.

BERT Igbo

Davlan · Nigeria

Open

Params: 110M
Context: 512
Min VRAM: 2 GB
GGUF: No

Fine-tuned for Igbo NER and text classification.

BERT Amharic

Davlan · Ethiopia

Open

Params: 110M
Context: 512
Min VRAM: 2 GB
GGUF: No

Fine-tuned for Amharic NER and text classification.

Naija Twitter Sentiment AfriBERTa

Davlan · Nigeria

Open

Params: 280M
Context: 512
Min VRAM: 4 GB
GGUF: No

Fine-tuned for Nigerian Twitter sentiment. Handles Pidgin.

XLM-R Yoruba (Davlan)

Davlan · Nigeria

Open

Params: 270M
Context: 8,192
Min VRAM: 16 GB
GGUF: No

Large context. Multilingual generation. Yoruba support.

Abena Base Akuapem Twi

GhanaNLP · Ghana

Open

Params: 110M
Context: 512
Min VRAM: 2 GB
GGUF: No

First Twi language model. NER and classification.

Wolof BERT

Abdouaziiz (Cheikh Anta Diop community) · Senegal

Open

Params: 110M
Context: 512
Min VRAM: 2 GB
GGUF: No

First Wolof language model. NER and classification.

MARBERT

UBC-NLP · Morocco

Open

Params: 160M
Context: 512
Min VRAM: 2 GB
GGUF: No

Handles Maghrebi dialects. Large Arabic corpus.

Tunisian Arabic RoBERTa

betteib (Tunisian community) · Tunisia

Open

Params: 120M
Context: 512
Min VRAM: 2 GB
GGUF: No

Specialized for Tunisian dialect. Sentiment analysis.

Al-Atlas-LLM 2B (Morocco Darija)

BounharAbdelaziz (Atlasia community) · Morocco

Open

Params: 2B
Context: 512
Min VRAM: 4 GB
GGUF: No

Multilingual. Handles Darija. Large context.

Nucleotide Transformer 500M

InstaDeep · Tunisia

Open

Params: 500M
Context: 1,000
Min VRAM: 8 GB
GGUF: No

Biotech focus. Large context. Genomic understanding.

EthioLLM-l-250K

EthioNLP · Ethiopia

Open

Params: 130M
Context: 1,024
Min VRAM: 2 GB
GGUF: No

First multilingual LLM covering 5 Ethiopian languages including Geez-script Tigrinya.

Amharic-LLAMA-MT

EthioNLP · Ethiopia

Open

Params: 7B
Context: 4,096
Min VRAM: 10 GB
GGUF: No

Continual-pretrained Llama 2 7B on Amharic; fluent generation in Geez script.

GalsenAI xTTS-v2 Wolof

GalsenAI (Senegal community) · Senegal

Open

Params: 500M
Context: 2,048
Min VRAM: 4 GB
GGUF: No

First public Wolof TTS with voice cloning, runs on consumer GPU.

XLM-R Large MasakhaNER

Davlan / Masakhane · Nigeria

Open

Params: 550M
Context: 512
Min VRAM: 2.5 GB
GGUF: No

Best public NER for 10 African languages; Davlan and Masakhane gold standard.

MARBERTv2

UBC-NLP (Canada / Tunisia) · Egypt-Africa / Maghreb

Open

Params: 160M
Context: 512
Min VRAM: 2 GB
GGUF: No

Strongest open encoder for dialectal Arabic including Egyptian, Tunisian, Moroccan.

Harrier-270M-Amharic

kiyam (Ethiopia community) · Ethiopia

Open

Params: 270M
Context: 8,192
Min VRAM: 1 GB
GGUF: Yes

Tiny 270M params yet generates coherent Amharic; runs on phones and Raspberry Pi.

XLM-R Swahili (Davlan)

Davlan · Nigeria

Open

Params: 270M
Context: 512
Min VRAM: 2 GB
GGUF: No

Strongest open Swahili encoder for classification and NER.

Hardware buying realities (by country)

Customs duties, power grid, used markets and shipping routes.

Import route: Air freight via Murtala Muhammed International Airport (LOS) or sea via Apapa Port, Lagos; couriers like DHL/FedEx handle customs clearance.
Vat customs: 5% import duty on computers, 7.5% VAT on total CIF value, plus 1% Comprehensive Import Supervision Scheme (CISS) levy.
Power grid reality: Grid unreliable with frequent outages; average 4-8 hours of power per day in major cities, necessitating UPS and generators.
Where to buy: Computer Village, Ikeja, Lagos for new/used parts; Slot Nigeria and Pointek for retail; Jiji.ng for peer-to-peer used gear.
Used market: Jiji.ng and Computer Village vendors offer used GPUs (e.g., GTX 1080 Ti, RTX 3060) at 30-50% below new prices, but verify authenticity.
Shipping cost estimate from us: FedEx/DHL shipping for a GPU (2kg) costs $80-$150; sea freight for larger items (10kg) $200-$400, 4-8 weeks.
Solar homelab relevance: High: solar + battery (e.g., 5kW system) is cost-effective for 24/7 AI training; many homelabs use 48V DC setups.
Gpu availability: Scarce for high-end (RTX 4090); mid-range (RTX 3060, 4060) available via import or used market; expect 20-40% markup.
Notes: Avoid shipping lithium batteries via air; use sea freight for powerwalls. Naira volatility affects pricing; pay in USD if possible.

Undersea cables internet quality

Africa connected by 16 submarine cables (e.g., SAT-3, WACS, ACE, 2Africa); internet quality varies: South Africa (50ms latency), Nigeria (80ms), landlocked countries (150ms+).

Brain drain to west

Top AI talent emigrates to US/EU for better pay and infrastructure; initiatives like Deep Learning Indaba aim to retain talent locally.

Africa genai fund world bank

World Bank's $5B Africa GenAI Fund (2024) supports local AI startups and compute infrastructure; disbursement slow.

Deep learning indaba

Annual pan-African ML conference (since 2017) held in different African cities; provides training, networking, and GPU access via sponsors.

Masakhane collective role

Grassroots NLP research network focused on African languages; provides datasets, compute credits (via Google Cloud), and mentorship.

Smart africa alliance

AU-led initiative to harmonize digital policies and promote local data centers; aims to reduce import barriers for tech hardware.

Notable vendors

Lelapa AI (ZA)Masakhane NLP (continental)Davlan / David Adelani (Nigeria, Saarland Univ.)EthioNLP (Ethiopia)Ghana-NLPGalsenAI (Senegal)Atlasia / BounharAbdelaziz (Morocco)betteib (Tunisia community)Castorini lab (Nigerian-led)InstaDeep (Cape Town / Dakar / Tunis)UBC-NLP (Maghrebi Arabic NLP)

Run benchmarks

llama.cpp, MLPerf, and inference leaderboards.

VRAM calculator

Will this model fit on your GPU?

Hardware guides

Local LLM setup, homelab, mini-PC builds.

Methodology

How we benchmark, verify ASINs, and source data.

Beginner's guide

New to local AI? Start here.

Sourcing & commerce disclosure

Model metadata on this page is compiled from each vendor's public HuggingFace repo, model card, or research paper at time of indexing. Performance figures (KMMLU, MMLU, HAERAE, AraBench, IndicGEM, etc.) link directly to source benchmarks where available; missing or partial values are shown as ", ".

Outbound hardware links (GPUs, mini-PCs, accelerators) are Amazon Associate links using store ID fredoline-20. As an Amazon Associate we earn from qualifying purchases. Prices and availability shown elsewhere on the site are snapshots, not live quotes — verify on Amazon before purchasing. See About / disclosures for the full policy.