Build Anywhere.
Run Locally.
OribAI integrates with the tools you already use — from Hugging Face Transformers to llama.cpp. One model, every environment.
Integrations
Seamless compatibility with leading deep learning frameworks and local inference runtimes.
Hugging Face
OribAI-14B is hosted on Hugging Face Hub. Load directly with transformers using AutoModelForCausalLM.from_pretrained().
Transformers
Full HuggingFace Transformers compatibility. Use AutoTokenizer and AutoModelForCausalLM with standard chat templates out of the box.
Unsloth
Optimized 4-bit inference and fine-tuning via Unsloth's FastLanguageModel. Significantly reduces VRAM requirements for 14B inference.
llama.cpp
Run the Q4_K_M GGUF on CPU-only hardware. Download the quantized file from Hugging Face and run locally with zero cloud dependency.
Ollama
Create a local Modelfile pointing to the GGUF and run OribAI with a single ollama run command. Zero configuration for local development.
LangChain
Wrap OribAI in LangChain's HuggingFacePipeline to use it inside chains, agents, and retrieval-augmented generation pipelines.
Hardware Compatibility
The Q4_K_M GGUF variant runs on CPU-only hardware. The full BF16 model requires a GPU with sufficient VRAM. Designed for the realities of African infrastructure.
Apple M-series (16 GB)
GGUF Q4 · ~15 tok/s
NVIDIA RTX 3090 (24 GB)
BF16 full · ~55 tok/s
NVIDIA A100 (80 GB)
BF16 full · ~180 tok/s
CPU-only (16 GB RAM)
GGUF Q4 · ~3 tok/s
Unsloth + GPU (≥16 GB)
4-bit NF4 · optimized
Join the Community
The frontier of African AI is shaped by the people building with it.
Community Forum
Ask questions, share projects, and connect with developers building on OribAI.
GitHub Repository
Contribute training scripts, evaluation benchmarks, and dialect data for v2.
Tutorials & Guides
Step-by-step guides for common use cases: dialogue, translation, and factual Q&A.