enver / ayncoding-qwen3-8b-slim

🏛️ AynCoding-Qwen3-8B-Slim: Sovereign Epistemic Classical Arabic Logic (Manṭiq) & Hierarchical MoE Engine

AynCoding-Qwen3-8B-Slim ( ayncoding-qwen3-8b-slim ) is an elite, sovereign software synthesis and architectural reasoning engine based on a knowledge-purged, layer-sliced Qwen3-8B architecture (2.8 GB GGUF) , aligned with Classical Arabic Logic ( Manṭiq ) , Morphological Root Lexicography ( Ishtiqāq ) , and Hierarchical Symbolic-Neural MoE (H-MoE) Acceleration .

Unlike standard models trained on unstructured, noisy GitHub repositories, AynCoding-Qwen3-8B-Slim actively purges illogical pre-trained habits (vague identifiers, circular dependencies, infinite loops, silent exceptions) and natively engages an epistemic Chain-of-Thought reasoning block ( <ayn_mantiq> ... </ayn_mantiq> ) prior to code generation.

🚀 Key Innovations

1. Hierarchical Symbolic-Neural MoE (H-MoE) Architecture

• Fast Speculative Drafting (1.5B): Generates high-frequency syntax at 60–70 tokens/sec.
• Deterministic Logic Gating (0.01s): Evaluates AST and 5 Classical Pillars in microseconds without GPU/CPU memory overhead.
• Targeted Epistemic Refinement (8B-Slim): Invoked only on demand to surgically patch complex logic nodes.
• Result: 27x faster generation latency (~13 seconds vs 355 seconds on CPU) while retaining 8B-grade reasoning.

2. Epistemic Knowledge Purge & Unlearning

Raw foundation models inherit noisy anti-patterns from the open web. We applied systematic unlearning and structural pruning:

3. Layer Slicing & Vocabulary Pruning

• Pruned 12 intermediate factual trivia layers (layers 16–27), preserving core syntax layers (1–15) and upper Ghazalian reasoning layers (28–35).
• Pruned 120,000+ non-programming multilingual tokens, reducing memory traffic by 42% (5.2 GB $ ightarrow$ 2.8 GB GGUF).

🌟 Governing Classical Authorities & Root Primitives

AynCoding is grounded in seven classical Arabic masterpieces:

1. Abū Ḥāmid al-Ghazālī ( Miʿyār al-ʿIlm & Miḥakk al-Naẓar ): Epistemic Definition by Essence and Fallacy Elimination.
2. Fakhr al-Dīn al-Rāzī ( Al-Mulakhkhaṣ fī al-Ḥikmah wa al-Manṭiq ): Law of Non-Contradiction and Syllogisms.
3. Al-Khalīl ibn Aḥmad al-Farāhīdī ( Kitāb al-ʿAyn ): Tri-consonantal Primitive Decomposition ( قفل , حفظ , نقل , عقد , حسب , حكم ).
4. Ibn Manẓūr ( Lisān al-ʿArab ): Exhaustive Error Taxonomy and Lifecycle State Coverage.
5. Al-Rāghib al-Iṣfahānī ( Al-Mufradāt fī Gharīb al-Qurʾān ): Teleological Purity and Intentional Identifiers.
6. Al-Zamakhsharī ( Asās al-Balāghah ): Concrete Reality ( Ḥaqīqah ) over Leaky Metaphors ( Majāz ).
7. Sībawayh ( Al-Kitāb ): Syntactic Governance ( Al-ʿĀmil wa al-Maʿmūl ), Static Typing, and AST Validity.

📊 Benchmark Results

🚀 Quickstart & Usage

1. Run with Ollama

# Create and run model from Modelfile
ollama create ayncoding-qwen3-8b-slim -f Modelfile
ollama run ayncoding-qwen3-8b-slim "Implement a thread-safe Monotonic Ring Buffer in Python"

2. Python SDK / AynCodingEngine

from core.coding_engine import AynCodingEngine

engine = AynCodingEngine(provider="ollama", model="ayncoding-qwen3-8b-slim")
result = engine.synthesize(
    prompt="Implement a Content-Addressed Immutable Store with SHA-256 integrity checks",
    language="python"
)

print(result["mantiq_reasoning"])
print(result["code"])

3. Accelerated Speculative Hybrid Execution

from core.speculative_engine import AynSpeculativeEngine

spec_engine = AynSpeculativeEngine(
    draft_model="ayncoding-model",
    verifier_model="ayncoding-qwen3-8b-slim"
)

# Generates 100% Grade A+ code in ~3-13 seconds on CPU
res = spec_engine.synthesize_accelerated(
    prompt="Implement a thread-safe Token Bucket Rate Limiter with bounded replenishment"
)
print(f"Generated in {res['duration_seconds']}s with Grade {res['epistemic_grade']}")
print(res["code"])