About the challenge
Africa is not excluded from the language model revolution. The bottleneck is no longer research or raw model availability - it is access economics. Cloud-hosted LLMs require API fees, stable fiber, and sustained electricity. For a university student in Lagos, an extension officer in Arusha, or a small-business owner in Dakar, these are not minor frictions - they are blockers.
Africa Deep Tech Challenge 2026 (inspired by the Africa AI XPrize) is an engineering-first competition to make useful language-model applications run well on the computers Africans already own: mid- and low-end commodity laptops. We are not building for exotic edge silicon. We are building for the 8 GB laptop with integrated graphics - the machine sitting on millions of desks, in classrooms, clinics, and corner shops across the continent.
You could be the one to democratize access to low cost inference for the continent.
This is an applied systems engineering contest. Hitting genuinely useful performance on commodity hardware requires optimization across the full stack: model selection and fine-tuning, quantization and compilation, memory and cache management, retrieval-augmented generation over local corpora, and application UX that makes such a model feel responsive.
Get started
-
Register on DevPost for the Africa Deep Tech Challenge 2026
- Form or join a team of 1 to 3 people and pick your problem domain
- Clone the submission template
- Test your model locally by cloning the model profiler. Ensure your model and submission formats are perfectly compatible with the ADTC evaluation pipeline. The local profiler allows you to run latency, throughput, memory, and CPU checks directly on your target hardware.
- Check out the rules and submission requirements, then start building!
Requirements
What to Build
Participants build a working, end-to-end, on-device Language Model that runs without cloud dependencies on the ADTC Standard Laptop (defined below). The model must address one of the published problem domains
Each team selects one primary domain. These are the ADTC standardized benchmarking domains; validation sets are provided for each.
-
Math & Scientific Reasoning - problem solving, proof assistance, scientific question-answering, and quantitative reasoning tasks.
-
Healthcare & Medical - clinical information, medical Q&A, triage support, and patient education.
-
Agriculture - crop, livestock, weather, and market advisory for farmers and extension officers.
-
Creative Writing - story generation, editing assistance, poetry, and narrative support across languages.
-
Coding Assistants - code generation, debugging help, and programming tutoring across common languages.
-
Corporate / Enterprise - knowledge-work productivity: summarization, drafting, and analysis for small and medium enterprises.
-
Autonomous AI Agents - local orchestration, task automation, and privacy-focused workflow management using decentralized tools and messaging interfaces.
All submissions are evaluated against a single, published reference hardware profile. This eliminates fragmentation and gives every participant a target they can design for. Participants may develop on any hardware, but final benchmarks and audits are reported against the Standard Laptop profile.
|
Component |
ADTC Standard Laptop |
|---|---|
|
CPU |
Intel Core i5 10th–12th gen OR AMD Ryzen 5 3000–5000 (x86-64) |
|
RAM |
8 GB DDR4 |
|
Graphics |
Integrated only (Intel UHD / Iris Xe or AMD Radeon integrated). No discrete GPU. |
|
Storage |
256 GB SSD |
|
OS (reference) |
Ubuntu 22.04 LTS |
|
Representative Price |
$400–$500 new / $150–$250 refurbished |
What to Submit
- Open Source Github repo that leverages the approved ADTC 2026 Report Template
- A comprehensive project report including:
- Problem definition and context
- Identified constraints (e.g. power, data, compute, connectivity)
- Documentation of design alternatives and final decisions
- Tools used and why they were chosen
- Performance tests and benchmarks
- Screenshots or short videos showing your build in action
- A short video (max 2 minutes) explaining your solution and development journey
- Updated repo, documentation and video (if part of semi-final or final round)
Leaderboard Scoring
Stotal = 0.50⋅Sacc+0.30⋅Sperf+0.20⋅Seff−Pthermal
|
Component |
Weight |
Formula |
Notes |
|
Sacc |
50% |
Weighted average of model response scored between 0 and 100 by a Judge. |
Weighted combination of automated benchmark scores and qualitative assessment of model prompt responses by the judge panel. |
|
Sperf |
30% |
100 × (TPSact ÷ TPSmax) |
TPS_REFERENCE = 15.0 provisional |
|
Seff |
20% |
Seff = 100 × ((7 GB − Peak RAM) ÷ 7 GB) |
Rewards lower RAM usage. The less memory consumed relative to the 7 GB budget, the higher the score. Peak RAM = 7 GB |
|
Pthermal |
-10 points |
-10 if throttled or temp > 85°C |
Else 0 |
Prizes
Grand Prize
Second Place
3rd Place
Best African Use Case
Finalist Stipends
$250 (In GPU Credits) for up to 10 finalists
Semifinalist Stipends
$50 (In GPU Credits) for up to 20 semi finalists
Devpost Achievements
Submitting to this hackathon could earn you:
Judges
Oji Udezue
Yannick Djoumbou Feunang
Oluwatobi Oyinlola
Christine Abernathy
Omoju Miller
Peter Ing
Mbangula Lameck Amugongo
Ola Fadiran
Judging Criteria
-
Model Accuracy & Quality
A combination of multiple-choice benchmarks and qualitative evaluations that includes accuracy of prompts, quality of documentation -
Model Throughput Performance
Evaluated relative to the maximum observed tokens per second -
Model Efficiency
Rewards lower RAM utilization profiles relative to the maximum memory budget -
African Use Case Bonus
Up to 10 extra points awarded for how applicable the model is to a real African use case -
Hardware & Thermal Penalties
10 points deducted if core/package temperature exceeds 85∘C or if thermal throttling is flagged. OOM or sandbox execution crash results in disqualification
Questions? Email the hackathon manager
Invite others to compete
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.