Akhet-Ra1 — Model Card & Documentation

A Local-First, Arabic-Aware Language Model by Alwkala Digital Labs

Part of the AkhetRa Ecosystem (Ra: Intelligence · Akhet: Governance)

Status: Release v0.1.0 · Date: 2026-10-23
Base Model: Qwen/Qwen3.5-4B
Fine-tuned & Quantized by: Alwkala Digital Labs
Model Type: Large Language Model (LLM) · Decoder-only causal
Deployment Focus: Local inference, Private infrastructure, and Self-Hosted AI
Active Distribution: Ra-1-4B-Q4_K_M.gguf (2.52 GB)
Hugging Face Repository: alwkala/Akhet-Ra1

---

ملخص تنفيذي بالعربية (Executive Summary in Arabic)

«آخت-رع 1» (Akhet-Ra1) هو نموذج ذكاء اصطناعي محلي مُحسَّن بالضبط الدقيق (Fine-tuned) استناداً إلى نموذج Qwen3.5 Base، طورته Alwkala Digital Labs ليكون طبقة الذكاء (رع / Ra) داخل منظومة آخت-رع (AkhetRa). يركز النموذج على الاستدلال المحلي والخصوصية التامة للبيانات، والتعامل الذكي مع اللغتين العربية والإنجليزية، ودعم سيناريوهات الوكلاء الذكيين (Agents) والأدوات الخاضعة للحوكمة الصارمة عبر بيئة التشغيل (آخت / Akhet).

تنبيه توثيقي: هذه البطاقة تمثل المواصفات المرجعية ومعايير التقييم للنموذج. تخضع حقول الأداء والقياسات الحسابية لنتائج الاختبارات المعيارية.

---

1. Overview & Identity

1.1 Overview

Akhet-Ra1 serves as the primary local intelligence layer within the AkhetRa architecture:

• Bringing AI capabilities directly to the edge and local hardware.
• Enforcing full organizational control over infrastructure and confidential data.
• Providing an integration-ready foundation for tool-calling and autonomous agents governed by the Akhet platform.

1.2 Model Identity Specification

Field	Value / Specification
Model Name	Akhet-Ra1 (GGUF Identifier: Akhet-Ra1-4B / Alias: Akhet-Ra-1)
Model Family	Akhet-Ra
Organization / Developer	Alwkala Digital Labs (alwkala.com)
Base Model	Qwen3.5 4B (Qwen/Qwen3.5-4B)
Model Type	Hybrid SSM + Self-Attention Causal Language Model (Qwen3.5 Architecture)
Parameter Count	4,205,751,296 (~ 4.2 Billion Parameters)
Vocabulary Size	248,320 tokens
Embedding Dimension	2,560
Hidden Layers / Blocks	32 Blocks (Hybrid SSM Mamba-style layers + Full Attention every 4 layers)
Native Trained Context	262,144 tokens (256K native training context)
Active Runtime Context	32,768 tokens (32K allocated in local server)
Reasoning Mode	Native Chain-of-Thought (<think>...</think> / reasoning_content)
Tool Calling Format	Native XML/JSON Tool Calling (<tool_call>, <tool_response>)
Base License	Apache-2.0 (general.license: apache-2.0)
Quantization Format	GGUF Q4_K_M (File: Ra-1-4B-Q4_K_M.gguf, Size: 2.52 GB)

---

2. Design Philosophy

1. Local-First: Compute resides on user hardware close to private data.
2. Arabic-Aware: Deep comprehension of Modern Standard Arabic, dialectal nuances (including Egyptian), technical terms, and code-switching.
3. Integration-Ready: Seamless coupling with runtime hooks, development tools, and agent loops.
4. Governance-Compatible: Designed to operate inside sandboxed runtimes enforcing strict tool access policies.
5. Model Agnostic Architecture: Operates independently while remaining interchangeable with external models under AkhetRa policies.

---

3. The AkhetRa Ecosystem Architecture

                  ┌──────────────────────────────────────────────┐
                  │                   AKHETRA                    │
                  │            The Integrated Ecosystem         │
                  └──────────────────────┬───────────────────────┘
                                         │
                 ┌───────────────────────┴───────────────────────┐
                 │                                               │
                 ▼                                               ▼
   ┌───────────────────────────┐                   ┌───────────────────────────┐
   │            RA             │                   │           AKHET           │
   │    The Intelligence Layer │                   │ The Runtime & Governance  │
   ├───────────────────────────┤                   ├───────────────────────────┤
   │ • Akhet-Ra1 LLM (Local)   │                   │ • Context & Memory Budget │
   │ • Reasoning & Generation  │◄─────────────────►│ • Tool Permissions       │
   │ • Arabic & English Logic  │  Strict Contract  │ • Sandboxing & Approvals  │
   │ • Tool Argument Synthesis │                   │ • Audit Trail & Logs      │
   └───────────────────────────┘                   └───────────────────────────┘

The Golden Principle: Ra thinks, Akhet governs.
The model produces proposals, reasoning, and tool call invocations; the Akhet runtime enforces security boundaries, schema validation, human approval, and telemetry.

---

4. Intended & Out-of-Scope Use Cases

4.1 Six Core Agentic Use Cases

1. Local Desktop Assistant: Fast, private desktop and workflow management without outbound network traffic. Handles file triage, local summarization, and task orchestration.
2. Triage & Routing Agent: Entity extraction and intent routing across bilingual Arabic/English queries, support tickets, and automation pipelines.
3. Data Parsing & JSON Structuring: Strict deterministic formatting of unstructured texts (invoices, receipts, logs) into schema-valid JSON for database ingestion.
4. Systems & Tool Orchestrator: Bridges conversational user intent with machine interfaces, executing structured function calls and system scripts within a governed sandbox.
5. RAG Query Generator: Decomposes complex user inquiries into targeted, dense search queries across vector stores and enterprise search engines.
6. Offline Educational & Language Drill: Real-time bilingual conversational drills, grammar feedback, and translation operating completely offline.

4.2 Operational Boundaries & Prohibited Uses

• Complex Multi-Step Logic Without Tools: Do not rely on unassisted causal generation for complex mathematical proofs or end-to-end full software architectures without external tools/interpreters.
• Unbounded Creative Prose: Avoiding long, rambling creative storytelling where context drift or hallucination degrades factual precision.
• Autonomous Unsupervised Critical Execution: Strictly prohibited from executing destructive shell/DB commands or taking medical/legal decisions without human approval.
• Uncited Sole Oracle: Not intended as an infallible single source of encyclopedic truth; should always be paired with RAG or structured tools for domain facts.

---

5. Deployment & Quick Start

5.1 Local Execution via llama.cpp (Bundled PowerShell Scripts)

The repository includes pre-configured launch scripts optimized for NVIDIA GPUs (6GB+ VRAM):

Interactive Terminal Dashboard (TUI)

# Starts the interactive live dashboard with GPU metrics, logs, and hotkeys
.\run-Ra-tui.ps1 -Port 8080 -ContextSize 32768

Headless Service Runner

# Launches llama-server as a background/daemon service
.\run-Ra-0.ps1

Raw llama-server Command

llama-server.exe \
  -m "A:\.lmstudio\Akhet\Ra-1\Ra-1-4B-Q4_K_M.gguf" \
  --alias "Akhet-Ra-1" \
  -ngl 99 \
  -c 32768 \
  -t 8 \
  --jinja \
  --flash-attn on \
  -ctk q4_0 \
  -ctv q4_0 \
  --batch-size 1024 \
  --ubatch-size 512 \
  --host 0.0.0.0 \
  --port 8080

5.2 Direct Execution via Hugging Face Hub

Thanks to Hugging Face Hub integration and the GGUF distribution, you can run or serve Akhet-Ra1 directly without manual file downloads:

Direct CLI Execution (llama-cli)

# Streams inference directly using weights hosted on Hugging Face
llama-cli -hf alwkala/Akhet-Ra1:Ra-1-4B-Q4_K_M.gguf -p "أنت مساعد ذكي محلي خاضع لمنظومة آخت-رع." -cnv

Direct Server Execution (llama-server)

# Launches an OpenAI-compatible server on port 8080 directly from the Hub
llama-server \
  --hf-repo alwkala/Akhet-Ra1 \
  --hf-file Ra-1-4B-Q4_K_M.gguf \
  --alias "Akhet-Ra-1" \
  -c 32768 \
  -ngl 99 \
  --host 0.0.0.0 \
  --port 8080

Direct Execution via Ollama

# Pull and run directly from Hugging Face Hub via Ollama
ollama run hf.co/alwkala/Akhet-Ra1:Ra-1-4B-Q4_K_M.gguf

Custom Ollama Modelfile

FROM ./Ra-1-4B-Q4_K_M.gguf
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>"""
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER num_ctx 32768

5.3 Downloading the GGUF Distribution Locally

You can download Ra-1-4B-Q4_K_M.gguf (2.52 GB) for offline local deployment:

Using Hugging Face CLI (hf):

hf download alwkala/Akhet-Ra1 Ra-1-4B-Q4_K_M.gguf --local-dir .

Using Python (huggingface_hub):

from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="alwkala/Akhet-Ra1",
    filename="Ra-1-4B-Q4_K_M.gguf",
    local_dir="."
)
print(f"Downloaded to: {model_path}")

Using LM Studio:

1. Open LM Studio.
2. Search for alwkala/Akhet-Ra1 in the model search bar.
3. Select Ra-1-4B-Q4_K_M.gguf and click Download.
4. In the System Prompt configuration, paste the system prompt from system_prompt.txt.
5. Load with GPU acceleration enabled (GPU Offload: Max).

5.4 Python Inference (Hugging Face Transformers / GGUF)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "alwkala/Akhet-Ra1"  # Or local path
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "system", "content": "أنت مساعد ذكي محلي خاضع لمنظومة آخت-رع."},
    {"role": "user", "content": "اشرح لي مبدأ الحوكمة الصارمة لنماذج الذكاء الاصطناعي."}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

5.5 OpenAI-Compatible Local Endpoint

Once running on http://localhost:8080, query via standard OpenAI clients:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="sk-no-key-required")

response = client.chat.completions.create(
    model="Akhet-Ra-1",
    messages=[
        {"role": "user", "content": "ما هي مميزات معمارية AkhetRa؟"}
    ]
)
print(response.choices[0].message.content)

---

6. Recommended Sampling Parameters

Parameter	Recommended Value	Description
Temperature	0.6 – 0.7	Balanced creativity and deterministic precision
Top-P	0.90	Nucleus sampling cutoff
Top-K	40	Limits token candidate pool
Repetition Penalty	1.10	Prevents cyclic repetitions in long context chains
Context Size	32,768 (32K)	Maximum trained / tested context window
KV Cache Quantization	q4_0 (-ctk q4_0 -ctv q4_0)	Halves KV cache VRAM footprint enabling 32K context on 6GB VRAM

---

7. Hardware & Resource Footprint

7.1 Distribution Footprint Overview

Distribution Format	File Size	Minimum VRAM / RAM	Tested Hardware	Full GPU Offload	Context Tested
GGUF Q4_K_M	2.52 GB	6 GB VRAM	NVIDIA RTX 3060 / Laptop 6GB	Yes (99 layers)	32K tokens
GGUF Q8_0	~4.3 GB (TBD)	8 GB VRAM	NVIDIA RTX 3070 / 4060 Ti	Yes	32K tokens
BF16 SafeTensors	~8.0 GB (TBD)	12 GB VRAM / 16 GB RAM	NVIDIA RTX 3060 12GB / A10G	Yes	32K tokens

7.2 VRAM & KV Cache Scaling Profile (Ra-1-4B-Q4_K_M.gguf)

Calculated using Hugging Face hf-mem memory estimation formulas for 4.2B parameters with Q4_K_M weights (2.52 GB):

Context Length	Weights VRAM	KV Cache (q4_0)	KV Cache (f16)	Total VRAM (q4_0)	Target Hardware
4,096 (4K)	2.52 GB	~0.35 GB	~0.70 GB	~2.87 GB	4GB GPUs / Integrated (Apple M-series)
8,192 (8K)	2.52 GB	~0.70 GB	~1.40 GB	~3.22 GB	4GB – 6GB GPUs
16,384 (16K)	2.52 GB	~1.40 GB	~2.80 GB	~3.92 GB	6GB VRAM GPUs
32,768 (32K)	2.52 GB	~2.80 GB	~5.60 GB	~5.32 GB	6GB VRAM GPUs (with -ctk q4_0 -ctv q4_0)

---

8. Evaluation Framework & Methodology

All benchmark claims must be validated through uniform protocols against the upstream base checkpoint (Qwen3.5 Base):

Evaluation Dimension	Benchmark / Harness	Target Metric	Base	Akhet-Ra1	Δ
Arabic Knowledge	ArabicMMLU	Accuracy (%)	[Pending]	[Pending]	[TBD]
Egyptian & Regional Dialects	Internal Egyptian Dialect Suite	Comprehension & Fluency	[Pending]	[Pending]	[TBD]
Instruction Following	IFEval	Strict Prompt Adherence	[Pending]	[Pending]	[TBD]
Reasoning & Logic	GSM8K / ARC-C	Accuracy (%)	[Pending]	[Pending]	[TBD]
Code Generation	HumanEval / MBPP	Pass@1 (%)	[Pending]	[Pending]	[TBD]
Tool Calling & Schema	BFCL / JSON-Schema Benchmark	Schema Validity Rate	[Pending]	[Pending]	[TBD]
Context Needle	RULER / Needle In A Haystack	Retrieval Accuracy	[Pending]	[Pending]	[TBD]
Safety & Jailbreak	Refusal & Prompt Injection Suite	Robustness Score	[Pending]	[Pending]	[TBD]

---

9. Safety, Governance & Alignment

1. Defense in Depth: The model's safety refusals are backed by the Akhet runtime firewall.
2. Untrusted Data Isolation: All web content, external RAG chunks, and tool outputs are treated as untrusted strings and never directly interpreted as system prompts.
3. Execution Guardrails: Any write, network, or process spawning action requires explicit user permission or pre-authorized policy tokens.

---

10. Citation & Attribution

@misc{alwkala_akhetra1_2026,
  title        = {Akhet-Ra1: A Local-First, Arabic-Aware Fine-Tuned Language Model},
  author       = {Alwkala Digital Labs},
  year         = {2026},
  publisher    = {Alwkala},
  howpublished = {\url{https://alwkala.com}}
}

---

11. Organization

Alwkala Digital Labs
Website: alwkala.com
Advancing practical AI systems, local intelligence, and governed agent frameworks across the MENA region.