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 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.
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.
Participation - 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 and test your model locally - Apply for up to 5 hours of GPU credits courtesy of Udutech to train your model.
What to Build Participants build a working, end-to-end, on-device Language Model that runs without cloud dependencies on the ADTC Standard Laptop. The model must address one of the published problem domains: Math & Scientific Reasoning, Healthcare & Medical, Agriculture, Creative Writing, Coding Assistants, Corporate / Enterprise, or Autonomous AI Agents.
All submissions are evaluated against a single, published reference hardware profile. Participants may develop on any hardware, but final benchmarks and audits are reported against the Standard Laptop profile: - 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
What to Submit - Open Source Github repo that leverages the approved ADTC 2026 Report Template - A comprehensive project report including problem definition, constraints, design alternatives, tools used, performance tests, and screenshots/videos - A short video (max 2 minutes) explaining your solution and development journey
Leaderboard Scoring Stotal = 0.50⋅Sacc + 0.30⋅Sperf + 0.20⋅Seff − Pthermal - Sacc (50%): Weighted average of model response scored between 0 and 100 by a Judge. - Sperf (30%): 100 × (TPSact ÷ TPSmax) - Seff (20%): Rewards lower RAM usage. The less memory consumed relative to the 7 GB budget, the higher the score. - Pthermal (-10 points): -10 if throttled or temp > 85°C. Else 0.
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
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