---
license: apache-2.0
language:
- ar
pipeline_tag: text-generation
tags:
- gguf
- llama-cpp
- egyptian-arabic
- conversational
- cpu
base_model: Qwen/Qwen2.5-1.5B-Instruct
---

# Mini-NABD — lightweight Egyptian-Arabic chat model (GGUF)

Mini-NABD is a tiny conversational model that speaks Egyptian Arabic (Masri)
like a friend: casual, funny when natural, short answers. It is an **additional**
fast/local companion to the main NABD AI, not a replacement.

## Files

| file | what | size | RAM |
|---|---|---|---|
| `mini-nabd-q4_k_m.gguf` | recommended 4-bit | ~1.0 GB | ~2 GB |

Only the smallest practical file is published — no variant zoo.

## Run it (CPU, 4GB RAM, no GPU)

```bash
# llama.cpp server (OpenAI-compatible)
./llama-server -m mini-nabd-q4_k_m.gguf -c 2048 --port 11435

curl http://127.0.0.1:11435/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "mini-nabd",
  "messages": [
    {"role": "system", "content": "أنت نبض الصغير — صاحب مصري خفيف الدم، رد بالعامية مختصر."},
    {"role": "user", "content": "عامل ايه؟"}
  ]}'
```

Or Python: `pip install llama-cpp-python`, then `Llama("mini-nabd-q4_k_m.gguf", n_ctx=2048)`.

- Recommended context: **2048** (fits 4GB with room to spare).
- Startup: seconds. Speed on a normal laptop CPU: roughly 15–40 tok/s
  (measured numbers go in the release notes per build).
- Temp 0.7, max ~256 tokens — it is tuned for short replies.

## NABD integration + Tavily example

```js
const { autoProvider, needsSearch } = require('./router'); // providers/router
const { tavilySearch, formatForMini } = require('./tavily'); // providers/tavily

async function answer(userText) {
  let context = '';
  if (needsSearch(userText)) {           // router decides, not the tiny model
    const t = await tavilySearch(userText);   // TAVILY_API_KEY stays server-side
    context = formatForMini(userText, t) + '\n';
  }
  const sys = 'أنت نبض الصغير — صاحب مصري خفيف الدم، رد بالعامية مختصر.'
    + (context ? '\n' + context : '');
  return miniReply(sys, userText);        // POST to 127.0.0.1:11435, fallback main
}
```

## Training (v0 status: base + system prompt)

v0 ships `Qwen2.5-1.5B-Instruct` Q4_K_M unchanged — the Egyptian personality
comes from the system prompt in `SYSTEM.md`, not from weights yet. Measured
gaps on the raw base (Arabizi loops, MSA-style refusals of light slang) are
documented and are exactly what v1 fixes.

v1 roadmap (ready to run, see `training/`): LoRA r=16 on 60 hand-written
Egyptian pairs (greetings, slang, Arabizi, humor, live-info refusals) on a
Colab T4, merge + Q4_K_M export replacing this file. Tool routing stays
external, so weights will contain zero tool tokens.

## Limitations

- Small talk + simple explanations only — hard reasoning, long code, and
  vision go to the main AI.
- Knowledge cutoff = base model; live facts need Tavily via the router.
- Egyptian dialect is a style layer over Qwen's Arabic — MSA slips through
  on technical topics.

## License

Apache-2.0 (inherits Qwen2.5 + Llama.cpp ecosystem licensing for the base).