ABUZ8 is a one-person lab. These are its brains, published with their scores, failures and training runs. The goal is the SUPERBRAIN fusion: Arabic-root reasoning, tool use and memory, fast local inference and new architecture in one open system.
Labels are honest: open weights means you can download and run it today; research means code and logs, not a finished model.
Open weights · Apache-2.0A 9B Arabic-first model built on QCRI/Fanar-1-9B, trained to reason through Arabic morphology: find a word's root, build its forms, and do Arabic math. Quantized GGUF builds run in llama.cpp, Ollama and LM Studio.
Open adapters · Apache-2.0The agent brain of QADIR_CORE: LoRA adapters for Qwen2.5-7B-Instruct-1M, trained to call tools reliably and keep a consistent persona across long sessions.
ResearchAn architecture experiment: geometric “triangle” organs added to Qwen3-0.6B to test whether structure lets a tiny model do agent tool tasks. Published as an honest log: the current result is negative and no weights are released.
ResearchResearch code for running large Mixture-of-Experts models on consumer GPUs: layer and expert paging, ternary and low-bit CUDA kernels, dequantization and proof scripts. An engine, not a model.
Every graded training run, the proof repo, and the Arabic data packets the brains learn from.
Run SUPERBRAIN-1 locally: download a GGUF from its Files tab and load it in llama.cpp, Ollama or LM Studio. Base-model licenses (Fanar, Qwen) apply on top of ours; read each model card before commercial use.