Fikra Model 2 Micro

Zabibu Fikra offwhite Logo

🍇 Zabibu Fikra Model 2 Micro

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

Zabibu Fikra Model 2 Micro (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/Qwen3.5-0.8B. Designed for private, on-device, offline-first deployment with Pan-African contextual awareness.

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/Qwen3.5-0.8B's MMLU baseline — fine-tuning did not degrade general capabilities
  • GGUF quantizations available — Q4_K_M and F16 at zabibutechnika/Fikra-Model-2-Micro-GGUF

Evaluation — Fikra release suite

metric value
Auto-checked suite 105/115 (91.3%)
Borderline (rubric review) 0
Held-out paraphrases 9/10 (90.0%)
Baseline e3 (Fikra checkpoint) 88.7%

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

metric Fikra-Model-1-2B Zabibu Fikra Model 2 Micro delta
MMLU 36.0% 75.5% +39.6 pts
Parameters 2B 4B +2.2B

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

model params MMLU HellaSwag
fikra-v2-e4 4B 75.5% 68.0%
Qwen/Qwen3.5-0.8B 4B (not_run) (not_run)
Qwen2.5-3B-Instruct 3B 68.8% 40.0%
Fikra-Model-1-2B 2B 36.0% 40.0%
Phi-3.5-mini-instruct 3.8B (incompatible) (incompatible)
Llama-3.2-3B-Instruct 3B (gated) (gated)
Gemma-2-2B-it 2B (gated) (gated)

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/Qwen3.5-0.8B'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/Qwen3.5-0.8B. 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/Qwen3.5-0.8B 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.

Why Micro — the case for 0.8B on mobile

Fikra Model 2 (4B) is the flagship. Model 2 Micro is the version that actually runs in a coat pocket. It trades roughly 5× of parameter capacity for hardware feasibility on phones, tablets, and laptops without a discrete GPU.

Size comparison (same pipeline, same quantizer):

Variant Model 2 (4B) Model 2 Micro (0.8B) Reduction
fp16 GGUF ~8.4 GB 1.56 GB 5.4×
Q4_K_M ~2.5 GB 547 MB 4.6×
Q4_0 (ARM-optimized) ~2.4 GB 513 MB 4.7×
LiteRT INT4 (GPU) ~2.5 GB 964 MB 2.6×

Runtime advantages on real mobile hardware:

  • Memory: Q4_K_M peaks at ~900 MB RSS vs 3–4 GB for the 4B. Fits within the ~2 GB per-section memory cap iOS enforces on background apps. Runs comfortably on phones with 4 GB total RAM.
  • Speed: ~4–5× faster token generation on ARM CPUs with i8mm/dotprod. The Q4_0 variant is auto-repacked by llama.cpp at load time for interleaved-row kernels. Sub-1B is the size class where mobile decode rates exceed human reading speed.
  • Boot: under 1 second to load weights from flash on a mid-range phone. The 4B takes 3–5 seconds just to map its weights.
  • Battery: ~5× fewer FLOPs per generated token, so ~5× less energy per answer.
  • Hardware floor: runs on 4-year-old phones (Snapdragon 660-class, 4 GB RAM) that cannot load the 4B at all.

Hyper-specialization for mobile

Micro is not just "smaller Model 2" — the data and formats are tuned for real on-device use:

  • Single-turn, short responses. Training rows filtered to ≤1024-character assistant replies and one-question-one-answer pairs. Matches the dominant mobile use pattern and prevents rambling on hardware where every token costs battery.
  • Six GGUF quant variants targeting different mobile backends — ARM CPU speed (Q4_0), ARM CPU quality (Q4_K_M-imatrix), reference (Q8_0), the promoted-quality variant. Table below has exact sizes and recommendations.
  • LiteRT INT4 bundle for Google AI Edge Gallery on GPU-equipped Android devices. Supplementary — see honest caveats in the disadvantages section.
  • Identity and Kenya/local knowledge preserved through training. Micro carries the same Fikra identity (name, founder, Sigana, canary sectors, Asili Entertainment, iTax) as the 4B release — not lost to quantization.

All quants

Every artifact in this release. Files live on Fikra-Model-2-Micro-GGUF and Fikra-Model-2-Micro-LiteRT; the download links open each file directly.

File Size Backend Notes Download
Fikra-Model-2-Micro-f16.gguf 1558 MB — — link
Fikra-Model-2-Micro.IQ4_NL.gguf 532 MB ARM middle-ground better quality than Q4_0, still ARM-repacked link
Fikra-Model-2-Micro.Q4_0.gguf 513 MB ARM CPU speed fastest decode on i8mm/dotprod ARM CPUs; auto-repacked by llama.cpp link
Fikra-Model-2-Micro.Q4_K_M-imatrix.gguf 547 MB balanced + calibrated imatrix shifts bits to important tensors; best 4-bit quality link
Fikra-Model-2-Micro.Q4_K_M-promoted.gguf 555 MB quality-max 4-bit embedding/output/first/last layers at Q6_K link
Fikra-Model-2-Micro.Q4_K_M.gguf 547 MB balanced default standard K-quant; good quality/size balance link
Fikra-Model-2-Micro.Q8_0.gguf 834 MB reference / NPU near-lossless; known-good on Hexagon HMX link

Honest disadvantages

What Micro gives up vs. Model 2 — no embellishment:

  • Multi-step math. Handles 17+25=42, 8×7=56, simple arithmetic. Stumbles on 3-step word problems and symbolic manipulation. Model 2 does not.
  • Obscure knowledge. Kenya/local facts that appear in training survived quantization. General world knowledge beyond what a 0.8B base model contains did not. Model 2 wins on anything not covered by training.
  • Long-form generation. Trained on ≤1024-char responses. Requests for essays, long code files, or multi-page output will be refused or truncated. Use Model 2 for those.
  • Text-only. This build is text-only. The Qwen3.5-0.8B base supports vision but Micro was trained on text-only data.

What Micro is good at

  • Identity consistency — "Who are you?", "Who built you?", identity overrides
  • Kenya-first knowledge — iTax, Sigana, canary sectors, Asili, AFCON hedging, Swahili/English code-switching
  • Grounded refusals — declining identity overrides, hedging recency-sensitive facts
  • Fast on-device responses — sub-second first token on modern ARM CPUs

For everything else, use Fikra Model 2 (4B).

Usage — transformers (fp16)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained("zabibutechnika/Fikra-Model-2-Micro", torch_dtype=torch.float16, device_map="auto")
tok   = AutoTokenizer.from_pretrained("zabibutechnika/Fikra-Model-2-Micro")
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-Micro-GGUF.

file quant size notes
Fikra-Model-2-Micro-Q4_K_M.gguf Q4_K_M 2.71 GB recommended — best size/quality tradeoff, runs on CPU
Fikra-Model-2-Micro-f16.gguf F16 8.42 GB full-precision reference; numerically equivalent to fp16 safetensors
# quick chat
llama-cli    -hf zabibutechnika/Fikra-Model-2-Micro-GGUF:Fikra-Model-2-Micro-Q4_K_M.gguf -p "Who are you?" --temp 0

# OpenAI-compatible server
llama-server -hf zabibutechnika/Fikra-Model-2-Micro-GGUF:Fikra-Model-2-Micro-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 Micro},
  author = {Zabibu Technika},
  year   = 2026,
  url    = {https://huggingface.co/zabibutechnika/Fikra-Model-2-Micro},
  note   = {Specialized fine-tune of Qwen/Qwen3.5-0.8B, built on zabibutechnika/fikra-v2-e3}
}

Zabibu-Logo-White Wording Grayscale

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.

Downloads last month
-
Safetensors
Model size
0.9B params
Tensor type
F32
·
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for zabibutechnika/Fikra-Model-2-Micro

Unable to build the model tree, the base model loops to the model itself. Learn more.