Zabibu Fikra Model 2 (4B)

Zabibu Fikra Model 2 (4B)

Answers in Swahili and English from a 4B model made to run privately on your own device, offline. Suits on-device assistants that need African context, such as Kenyan data.

Made by zabibutechnika on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Zabibu Fikra Model 2 (4B) is a 4 billion parameter chat model from Zabibu Technika in Nairobi. It is a fine-tune of their fikra-v2-e3 checkpoint, which comes from Qwen3.5-4B. It is built for private, offline use on your own device, with Swahili and English and African context in mind.

Who it suits: People who want a small assistant that works without a network, and teams who need Swahili and English in one model.

Industry | Who uses it | How they use it | Example
Software | App developers | Build an assistant that runs on the user's device | A chat feature that answers in Swahili or English with no server call
Edge and air-gapped IT | Deployment teams | Run the model where the network is off by design | An office machine with no internet access
Local products | Product teams | Handle users who switch between Swahili and English mid-sentence | A support helper for a Kenyan app
Research | NLP researchers | Compare a small model on African context cases | Try the cases the card lists, such as culture and local knowledge

You can skip it if you need a large hosted model or a language outside Swahili and English, since the card only lists those two.

Where to use it: The Safetensors weights load in the transformers library, and the GGUF builds open in desktop runtimes.

• Python with the transformers library
• llama.cpp
• Ollama
• LM Studio
• Any other GGUF-compatible runtime on Windows, macOS or Linux

Use case | Who | What they get back
Private on-device assistant | App developers | Replies without any external call at inference
Swahili and English code-switching | Product teams | One model that follows both languages
Edge deployment | IT teams | A small model that runs on CPU in its Q4_K_M build

How to set it up: You need Python with pip to fetch the files, and disk space for the weights, about 8.4 GB in F16.

• Python with pip
• About 8.4 GB of disk space for the F16 weights, or about 2.1 GB at 4-bit as an estimate
• A runtime of your choice: transformers, llama.cpp, Ollama or LM Studio

1. Install the Hugging Face download tool.
2. Download the model files to your computer.
3. For the GGUF build, take it from the maker's companion GGUF repository instead.
4. Open the files in your runtime.

pip install -U huggingface_hub
hf download zabibutechnika/Fikra-Model-2-4B

How to use it: Load the weights in transformers, or open a GGUF build in llama.cpp, Ollama or LM Studio, then chat in Swahili or English. The card says the system prompt and deployment design assume no network at inference.

Tool | What it does
transformers | Loads the F16 Safetensors weights
llama.cpp, Ollama, LM Studio | Open the GGUF builds
Q4_K_M | The GGUF size that runs on CPU
F16 | The GGUF reference copy

Example: Ask it in Swahili who it is. The card's test suite has four Swahili identity cases and reports all four passed, so expect a Swahili reply about its identity.

For newcomers and experts: A language model reads your text and writes a reply. This one is small enough to run on your own computer, and the card says it needs no external calls to answer. A first small step is to download the GGUF build, open it in LM Studio and type a greeting in Swahili.

For experts, this is an instruction-tuned fine-tune of fikra-v2-e3, itself derived from Qwen3.5-4B. Most of the MMLU gain over the 2B Fikra model comes from the larger base, not from Fikra training, and the card says so. The Fikra-specific gain is measured on its own 115-case release suite, 2.6 points over the e3 checkpoint. The trade-off is size against evidence, because the public benchmark runs used only 10 samples per task.

Download the weights or a GGUF build
Open them in transformers or a GGUF runtime
Send a Swahili or English prompt
The model replies on your device, offline

Good to know: The release suite results below are the card's own numbers.

• Sample size: The MMLU and HellaSwag runs used a limit of 10 samples per task, so read them loosely.
• Not size-normalized: The card says the MMLU gain over the 2B model mostly comes from the bigger base.
• Claims it declines: The card does not call the model first, best in Africa or state of the art.

Title: Fikra release suite results
Auto-checked suite: 91.3%
Held-out paraphrases: 90.0%
Earlier e3 checkpoint: 88.7%

• License: Apache 2.0, and the card says the base model's license is set by Qwen3.5-4B.
• Cost to run: none named, the source lists no API key, account or paid service.
• Data it reaches: nothing beyond what you type in, since the design assumes no network at inference.
• Not for: claims of being the first African language model, the best in Africa or state of the art, which the card says it does not make.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 4.2B
• Weights: about 8.4 GB in F16; about 2.1 GB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 682
• Likes: 3
• License: apache-2.0
• Languages: en, sw
• Based on: zabibutechnika/fikra-v2-e3
• Quantized versions: 3
• Tags: transformers, safetensors, gguf, qwen3_5_text, text-generation, africa, kenya, swahili, on-device, offline-first, fikra, zabibu-technika

Links

• Model: https://huggingface.co/zabibutechnika/Fikra-Model-2-4B
• Files: https://huggingface.co/zabibutechnika/Fikra-Model-2-4B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download zabibutechnika/Fikra-Model-2-4B

Then follow the usage in its model card. Nothing in this kit runs by itself.