Research Preview Notice: AkiliCode-14B is MsingiAI's secondary product, positioned as a credible research preview in v1. Sauti is our primary commercial product. AkiliCode does not have a live interactive playground in this release (Phase 2 roadmap). Model weights and benchmark results are available on Hugging Face for independent research evaluation.
Model Card
Benchmark Results
| Benchmark | AkiliCode-14B | Base Model | Improvement |
|---|---|---|---|
| HumanEval pass@1 | 64.2% | 61.8% (Qwen2.5-Coder-14B) | +2.4% |
| MBPP pass@1 | 73.1% | 71.2% (Qwen2.5-Coder-14B) | +1.9% |
| Swahili-Code Comment Accuracy | 89.3% | N/A (novel benchmark) | New metric |
| East African API Convention Score | 76.5% | N/A (novel benchmark) | New metric |
* Swahili-Code Comment Accuracy and East African API Convention Score are novel benchmarks developed by MsingiAI. Full methodology available in our research publications.
Capabilities
Code generation in Python, JavaScript, TypeScript, and Go
Swahili and English code comment generation
East African fintech API integration examples
Structured reasoning on domain technical documentation
Zero-shot code translation between languages
Automated docstring and README generation
Phase 2 Roadmap
Interactive AkiliCode Playground
A live coding sandbox with pre-compiled Swahili-commented code examples and dry-run output is planned for Phase 2 of the MsingiAI revamp. Once Sauti is established as the platform's primary commercial proof point, AkiliCode will receive its own interactive interface.
Read the Research Publication →