AitroniXAI/Aitronix-A1.1S-GGUF

⚡ Aitronix A1.1S GGUF (AitroniX A1.1S)

---

📌 Overview

Aitronix A1.1S GGUF is the official high-precision quantized GGUF release of Aitronix A1.1S, an advanced 2.4 Billion parameter (2.4B) bilingual (English & Arabic) frontier-grade reasoning and agentic model.

Quantized in Q5_K_M (5-bit medium precision), Aitronix A1.1S preserves exceptional mathematical accuracy, strict AST function calling fidelity, and syntactic coherence that matches unquantized FP16 baselines while operating at ultra-fast speeds with a lightweight footprint (~1.81 GB) on desktop GPUs, laptops, Apple Silicon (M1/M2/M3/M4), smartphones, and edge devices.

---

🏆 Benchmark Evaluations & Comparative Analysis

Aitronix A1.1S demonstrates outstanding performance across official evaluation benchmarks, outperforming much larger open frontier models including Google's Gemma-4-E2B (5.1B), Gemma-4-E4B (8B), and Alibaba's Qwen3.5-2B (2.6B) and Qwen3.5-4B (4.7B).

Official Evaluation Highlights

Domain	Benchmark	Aitronix A1.1S (2.4B)	Gemma-4-E2B (5.1B)	Gemma-4-E4B (8B)	Qwen3.5-2B (2.6B)	Qwen3.5-4B (4.7B)
Tool Use	BFCL v4 (Strict AST Match)	68.8%	36.6%	47.0%	43.6%	56.8%
Tool Use	τ²-Bench Telecom	97.1%	20.8%	20.8%	69.0%	92.1%
Tool Use	τ³-Bench Banking	20.8	3.9	4.1	2.1	6.8
Code Reasoning	LiveCodeBench v6	69.1%	42.9%	53.9%	20.2%	56.4%
Code Reasoning	LCB-Pro 25Q2 (Easy)	68.0%	27.1%	45.8%	10.3%	58.3%
Math Reasoning	AIME 2025	86.5%	31.7%	37.1%	29.6%	78.8%
Math Reasoning	MATH-500	94.6%	85.4%	88.2%	85.8%	99.0%
Instruction	IFStruct	85.49%	64.85%	76.65%	36.25%	78.50%
Instruction	IFBench	66.3%	25.7%	28.3%	46.0%	59.0%
Instruction	Multi-IF	71.8%	40.3%	45.9%	57.1%	73.6%
STEM	AA Omniscience (Truthfulness)	-29.50	-74.47	-49.03	-54.30	-50.43
Agentic	GAIA Text-103	88.7%	30.1%	39.5%	47.9%	78.6%
Agentic	SWE-bench Verified	46.4%	2.0%	15.0%	5.0%	33.6%
Agentic	Claw-Gym	59.2%	31.3%	37.9%	25.5%	51.6%
Overall	Benchmark Average	53.9	24.6	31.2	28.0	51.1

• BFCL v4 (Berkeley Function-Calling Leaderboard): Evaluated across strict AST ground truth evaluation. Aitronix A1.1S achieves 68.8% strict AST match, beating Alibaba's Qwen3.5-4B (56.8%) and Google's Gemma-4-E4B (47.0%), while delivering flawless parameter schema validation and near-zero tool hallucination.
• Math & STEM: Scores 86.5% on AIME 2025 and 94.6% on MATH-500, delivering competition-grade mathematical reasoning at only 2.4B parameters.
• Instruction Following: Exceeds 85.49% on IFStruct and 71.8% on Multi-IF, demonstrating exceptional adherence to multi-constraint negative rules, format guidelines, and structured JSON schemas.
• Truthfulness & Calibration (AA Omniscience): Highest truthfulness score (-29.50) in its weight class, substantially reducing hallucinations compared to Google's Gemma-4-2B (-74.47).

---

📊 Model Specifications

Parameter	Value
Model Name	Aitronix A1.1S GGUF (AitroniX A1.1S)
Base Architecture	MiniCPM5-2B Dense Transformer with GQA & RoPE
Parameters	~2.4 Billion (2.4B)
Quantization Format	GGUF (Q5_K_M — 5-bit Medium Precision)
File Name	Aitronix-A1.1S-Q5_K_M.gguf
File Size	~1.81 GB (High-precision compact footprint)
Context Window	8,192 tokens (8K)
Target Platforms	Laptops, Desktop GPUs, Apple Silicon, Smartphones, Edge Hardware
Supported Languages	English (EN) & Arabic (AR)
Supervision	Eng. Ahmed Hafez
Lead AI Engineer	Mohamed Mostafa Helal
Organization	AitroniXAI

---

🚀 Key Features

• 💎 High-Precision Q5_K_M Quantization: Provides a superior accuracy-to-size balance over conventional 4-bit quants. Extra bits allocated to critical attention projections ensure minimal perplexity degradation.
• 🤖 Autonomous Agent & Tool Execution: 68.8% strict AST matching on BFCL v4 and 97.1% on τ²-Telecom. Exceptional capabilities in multi-step tool execution, parameter parsing, and API calling.
• 📐 Mathematical & Code Reasoning: 86.5% on AIME 2025 and 69.1% on LiveCodeBench v6, outperforming rival models twice its size.
• 🌍 Native Bilingual Fluency: Seamless performance across Arabic and English with natural phrasing, cultural nuances, and idiomatic expression.
• ⚡ Ultra-Low Latency & Memory Footprint: Runs smoothly with full GPU offload on low-VRAM GPUs (GTX 1650/1660, RTX 2060/3050/4060, Apple M-series, and mobile devices).

---

⚙️ Recommended Generation Settings

Parameter	Recommended Value	Description
Temperature	0.7 – 1.0	Balanced reasoning accuracy and natural fluency
Top-P	0.85 – 0.95	Nucleus sampling for coherent output
Repetition Penalty	1.05 – 1.10	Prevents token loops while keeping technical terminology intact
Max Tokens	4096 – 8192	Full context length for long agentic trajectories

---

🛠️ Prompt Format

Aitronix A1.1S utilizes the ChatML prompt template:

<|im_start|>system
You are Aitronix A1.1S, an intelligent AI assistant developed by the AitroniX Team under the supervision of Eng. Ahmed Hafez.<|im_end|>
<|im_start|>user
{user_query}<|im_end|>
<|im_start|>assistant
{response}<|im_end|>

---

💻 How to Use

1. LM Studio (Desktop)

1. Download Aitronix-A1.1S-Q5_K_M.gguf.
2. Move the file into your LM Studio models folder.
3. Select ChatML as the preset chat template.
4. Set Temperature to 0.8 and Context Length to 8192.
5. Start chatting!

---

2. llama.cpp / CLI

Run inference directly using llama-cli:

llama-cli -m "Aitronix-A1.1S-Q5_K_M.gguf" \
  -co -cnv \
  -p "<|im_start|>system\nYou are Aitronix A1.1S, an intelligent AI assistant developed by the AitroniX Team under the supervision of Eng. Ahmed Hafez.<|im_end|>\n<|im_start|>user\nHello! How can you assist me?<|im_end|>\n<|im_start|>assistant\n" \
  --temp 0.8 \
  --repeat-penalty 1.05 \
  -c 8192 \
  -ngl 99

---

3. llama-server (OpenAI API Compatible Server)

Launch the built-in server with Web UI and OpenAI endpoints:

llama-server -m "Aitronix-A1.1S-Q5_K_M.gguf" \
  --host 127.0.0.1 \
  --port 8082 \
  -c 8192 \
  -ngl 99 \
  -fa auto \
  --alias "aitronix,aitronix-a1.1s"

Access the interactive web UI at http://127.0.0.1:8082.

---

4. Python (openai Client)

Query the running server with the official OpenAI Python SDK:

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8082/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="aitronix-a1.1s",
    messages=[
        {"role": "system", "content": "You are Aitronix A1.1S, an intelligent AI assistant developed by the AitroniX Team under the supervision of Eng. Ahmed Hafez."},
        {"role": "user", "content": "Explain how agentic tool execution works in small language models."}
    ],
    temperature=0.8,
    max_tokens=2048
)

print(response.choices[0].message.content)

---

5. Python (llama-cpp-python)

Direct in-process execution:

from llama_cpp import Llama

llm = Llama(
    model_path="Aitronix-A1.1S-Q5_K_M.gguf",
    n_ctx=8192,
    n_gpu_layers=-1  # Offload all layers to GPU/VRAM
)

output = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "You are Aitronix A1.1S, an intelligent AI assistant developed by the AitroniX Team under the supervision of Eng. Ahmed Hafez."},
        {"role": "user", "content": "Hello! What are your key capabilities?"}
    ],
    temperature=0.8,
    repeat_penalty=1.05,
    max_tokens=1024
)

print(output["choices"][0]["message"]["content"])

---

6. Mobile Inference (iPhone / Android)

Because this model is only ~1.81 GB:

• iOS (iPhone / iPad): Import Aitronix-A1.1S-Q5_K_M.gguf into apps such as Private LLM, Enchanted, or PocketLLM.
• Android: Run with Termux and llama.cpp or ChatGD.

---

👥 Team & Supervision

Developed by the AitroniX Team • Under the supervision of Eng. Ahmed Hafez

• Supervision & Technical Guidance: Eng. Ahmed Hafez
• Lead AI Engineer: Mohamed Mostafa Helal
• Organization: AitroniXAI
• Website: Aitronix.cloud

---

📜 Citation

@misc{aitronix_a11s_gguf_2026,
  author = {AitroniX Team, Ahmed Hafez, and Mohamed Mostafa Helal},
  title = {Aitronix A1.1S GGUF: High-Precision Q5_K_M Bilingual Reasoning and Agentic LLM},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/AitroniXAI/Aitronix-A1.1S-GGUF}}
}