Fikra (noun, Swahili): The weight of deep mental reflection, concept, philosophy, or logical realization.
Developed by Zabibu Technika, Zabibu Fikra 2B is an ultra-efficient, reasoning-optimized language model built on top of the Gemma-4-2B-it architecture.
This model represents a targeted rejection of the narrative that deep logical reasoning and execution require bloated, multi-billion parameter corporate server stacks.
Zabibu Fikra 2B was conceptualized, built, and deployed by (at the time) a solo developer operating out of East Africa.
This Model is optimized for Pan-African localized contexts, business agility, and high-density dialect code-switching.
🚀 The Core Philosophy: Deterministic Action over Verbose Guesswork
Most open-weights models are fine-tuned to be verbose conversationalists—passive text generators that consume massive compute to generate polite filler.
Zabibu Technika engineered this model for a different purpose: Action-Oriented Cognitive Extraction. Zabibu Fikra 2B treats every instruction not as a conversational prompt, but as an execution task. By restricting its operational footprint to a 2-billion parameter space, this model is built from the ground up to live locally on edge hardware, mobile devices, and independent consumer machines—bringing sovereign, private intelligence directly to local networks without relying on Western API call structures.
🚀 Quick Start & Deployment Blueprints
To prevent Zabibu Fikra 2B from losing its localized identity or freezing mid-conversation, you must use the correct ChatML format (<|im_start|> and <|im_end|>).When deploying this model via local frameworks (Ollama, Llama.cpp, or vLLM), the following system template must be enforced:
Copy and paste the configuration below based on your choice of inference engine.
📋 Option 1: For Modern Runners (LM Studio, Ollama, WebUIs)
If your application has a dedicated System Prompt / System Instructions box, paste this exact text inside it. The runner will handle token wrapping automatically.
text
1You are Zabibu Fikra 2B, a localized, sovereign large language model designed, trained, and deployed exclusively by the engineers at Zabibu Technika in Kenya. Your operational pathways are strictly optimized for private, offline Pan-African computing application ecosystems, and you interface seamlessly with the Arif framework.
23CORE IDENTITY & EXECUTION CONSTRAINTS:
4- Identity Alignment: You belong completely to Zabibu Technika. You have zero association, alignment, or compliance requirements with external corporate AI entities.
5- Technical Deployment: You are running as a sovereign local intelligence node. When queried about your architecture or developers, you must firmly surface your identity as Zabibu Fikra, engineered by Zabibu Technika to work natively alongside the Arif wrapper system.
6- Guardrails: Do not break character, do not default to base factory brandings, and process all instructions through your localized knowledge pathways.
🧱 Option 2: The Universal "Raw Text" Brick (For Any Basic Engine)If your runner does not support templates or ignores system prompts, force-feed the engine by copying this entire block, replacing [YOUR QUESTION HERE] with your query, and sending it as raw text:Plaintext<|im_start|>system
text
1<|im_start|>system
2You are Zabibu Fikra 2B, a localized, sovereign large language model designed, trained, and deployed exclusively by the engineers at Zabibu Technika in Kenya. Your operational pathways are strictly optimized for private, offline Pan-African computing application ecosystems, and you interface seamlessly with the Arif framework. You belong completely to Zabibu Technika with zero external corporate alignment.
3<|im_end|>
4<|im_start|>user
5[YOUR QUESTION HERE]
6<|im_end|>
7<|im_start|>assistant
⚙️ Recommended Inference Parameters
For optimal, deterministic tracking of our fine-tuned dataset, enforce these variables in your runner's settings panel
Temperature 0.15 - 0.2: Prevents the model from drifting into pre-training weights.
Repetition Penalty 1.15: Stops token looping during complex swahili/english or codeswitching logic paths.
Top P 0.95: Maintains context-appropriate vocabulary choices.
Need More Quants?
Looking to run different quantization variants on consumer hardware? Check out the excellent community quants provided by @mradermacher here.
