🍇 Zabibu Fikra Model 2 - 4B

Fikra (noun, Swahili): The weight of deep mental reflection, concept, philosophy, or logical realization.

Zabibu Fikra Model 2 - 4B (internal: fikra-v2-e4) is a 4B-parameter instruction-tuned language model developed by Zabibu Technika, Nairobi, Kenya. It is a Fikra-specialized fine-tune built on zabibutechnika/fikra-v2-e3, which in turn derives from Qwen/Qwen3.5-4B. Designed for private, on-device, offline-first deployment with Pan-African contextual awareness.

• Weights (fp16, transformers): zabibutechnika/Fikra-Model-2-4B — this repo
• Weights (GGUF, llama.cpp): zabibutechnika/Fikra-Model-2-4B-GGUF — Q4_K_M + F16
• Size (fp16): 8.43 GB
• Parameters: 4,205,751,296 (4B)
• License: apache-2.0 (base license: see Qwen/Qwen3.5-4B)
• Contact: zabibutechnika+fikra@gmsil.com

Highlights

• identity_direct — 8/8 (100%)
• identity_swahili — 4/4 (100%)
• uhi_protocol — 5/5 (100%)
• safety_boundary — 3/3 (100%)
• longform — 2/2 (100%)
• tool_abi — 4/4 (100%)
• Swahili-native identity — identity_swahili full pass; code-switching supported
• On-device, offline-first — no external calls required at inference
• African-context eval — 115-case release suite across Kenyan data, canary-review, gaming, fintech, health, culture, local knowledge
• Preserves Qwen/Qwen3.5-4B's MMLU baseline — fine-tuning did not degrade general capabilities
• GGUF quantizations available — Q4_K_M and F16 at zabibutechnika/Fikra-Model-2-4B-GGUF

Evaluation — Fikra release suite

In-suite (91.3%) and held-out (90.0%) are statistically indistinguishable (binomial p>0.05) — no contamination signal on the release suite.

Comparison to Model 1

These numbers are not size-normalized. Most of the MMLU delta comes from upgrading the upstream base (2B → 4B), not from Fikra training alone. The Fikra-specific improvement is measured on the release suite (+2.6 pts over the e3 Fikra checkpoint) and on the identity/context categories above.

Comparison to contemporary small models

Run conditions: lm-evaluation-harness 0.4.13, tasks mmlu,hellaswag, limit 10 samples per task, batch auto, fp16, on Kaggle 2×T4. Raw JSONs in benchmarks/.

Position

Verifiable claims

• Fikra specialization built in Kenya by Zabibu Technika, Nairobi, Kenya — the fine-tuning data, the release eval suite, and the deployment design.
• Preserves Qwen/Qwen3.5-4B's general capability while adding Fikra-specific behaviors measured at 91.3% on the Fikra release suite.
• African-context release eval — 115-case suite across identity-Swahili, Kenya data, canary-review, pan-African culture, gaming, fintech, health, local knowledge.
• On-device, offline-first. System prompt and deployment assume no network at inference.
• GGUF available — Q4_K_M (runs on CPU) and F16 (reference).

Claims this card does NOT make

• 'First African language model' — false (Lelapa/InkubaLM, Jacaranda, Masakhane, EthioLLM exist, and all fine-tune existing bases).
• 'We built the base model' — the base is Qwen/Qwen3.5-4B. Fikra is the specialization.
• 'Best in Africa' — unmeasured.
• 'State of the art' — unmeasured on any standard benchmark.
• 'Outperforms [model]' without the row above showing the same task/limit and after caveats 1–4 are resolved.

Intended use

On-device assistants requiring private/offline operation, Swahili/English code-switching, identity-preserving personas, edge/air-gapped deployment. Use the GGUF builds for llama.cpp, Ollama, LM Studio, or any GGUF-compatible runtime.

Limitations

• No cross-language Swahili benchmark — the differentiator is not measured by MMLU/HellaSwag.
• No live information — recency is limited to training data.
• 4B model — will hallucinate on specialized domains; not a replacement for retrieval.
• Reasoning-tag leakage — model occasionally emits <think>...</think> blocks in visible output. Strip in post-processing.
• Inherited biases — Qwen/Qwen3.5-4B base model biases carry through; not audited here.
• Q4_K_M quantization not re-evaluated — expect 1–3 pt loss vs fp16 on short-form evals. F16 GGUF inherits fp16 numbers exactly.

Safety

No unresolved safety case in the auto-checked suite.

Not red-teamed by an external party.

Usage — transformers (fp16)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained("zabibutechnika/Fikra-Model-2-4B", torch_dtype=torch.float16, device_map="auto")
tok   = AutoTokenizer.from_pretrained("zabibutechnika/Fikra-Model-2-4B")
msgs  = [{"role":"user","content":"Who are you, and where were you built?"}]
p     = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
i     = tok(p, return_tensors="pt").to(model.device)
o     = model.generate(**i, max_new_tokens=400, do_sample=False, pad_token_id=tok.eos_token_id)
print(tok.decode(o[0][i["input_ids"].shape[1]:], skip_special_tokens=True))

GGUF / llama.cpp

Both Q4_K_M and F16 conversions are published at zabibutechnika/Fikra-Model-2-4B-GGUF.

# quick chat
llama-cli    -hf zabibutechnika/Fikra-Model-2-4B-GGUF:Fikra-Model-2-4B-Q4_K_M.gguf -p "Who are you?" --temp 0

# OpenAI-compatible server
llama-server -hf zabibutechnika/Fikra-Model-2-4B-GGUF:Fikra-Model-2-4B-Q4_K_M.gguf --port 8080 -c 8192 -ngl 99

Quantization caveat: the eval numbers above are fp16. Q4_K_M typically loses 1–3 pts on short-form evals; F16 GGUF inherits the fp16 numbers exactly. Q4_K_M has not been re-evaluated — treat it as within-1–3-pts-of-fp16, not as equivalent.

Conversion provenance (llama.cpp build recipe, --no-mtp requirement for Qwen3.5, tokenizer_class patch) is documented in the GGUF card.

Citation

@misc{zabibu-fikra-model-2,
  title  = {Zabibu Fikra Model 2 - 4B},
  author = {Zabibu Technika},
  year   = 2026,
  url    = {https://huggingface.co/zabibutechnika/Fikra-Model-2-4B},
  note   = {Specialized fine-tune of Qwen/Qwen3.5-4B, built on zabibutechnika/fikra-v2-e3}
}

Release date: 2026-11-06T01:26:03Z. Card generated from RELEASE.json, results_20261105_233152.json, results_20261105_234218.json, BENCHMARKS.md, and the GGUF repo — all published alongside these weights.