| --- |
| license: apache-2.0 |
| base_model: empero-ai/Qwen3.8-4B-Distill |
| tags: |
| - text-generation |
| - qwen3_5 |
| - distillation |
| - reasoning |
| - full-stack |
| - typescript |
| - react-router-v8 |
| - nestjs |
| - flutter |
| - unsloth |
| inference: false |
| model_type: qwen3_5 |
| --- |
| |
| # Qwen3.8 4B Distill by Empero-AI and finetuned by iWebRoot |
|
|
| ## Model Overview |
| This model is a highly specialized, fine-tuned variant of **[empero-ai/Qwen3.8-4B-Distill](https://huggingface.co/empero-ai/Qwen3.8-4B-Distill)**, optimized using the **Unsloth** framework for modern full-stack web and cross-platform mobile software development. |
|
|
| The base model features a full-parameter distillation of reasoning traces (Chain-of-Thought via `<think>...</think>` tags) from the frontier-scale **Qwen3.8 2.4T A95B** teacher model developed by **Empero-AI**. This grants the student model frontier-level planning, logic, and self-correction behaviors while maintaining a highly compact 4-billion parameter footprint tailored for local consumer hardware. |
|
|
| ## 🔗 Repository Links |
| * **Safetensors Version (9.3GB Heavy Build):** [https://huggingface.co/iBotIA/Qwen3.8-4B-Empero-AI-FullStack](https://huggingface.co/iBotIA/Qwen3.8-4B-Empero-AI-FullStack) |
| * **GGUF Version (3.5GB Optimized Quantization):** [https://huggingface.co/iBotIA/Qwen3.8-4B-Empero-AI-FullStack-GGUF](https://huggingface.co/iBotIA/Qwen3.8-4B-Empero-AI-FullStack-GGUF) |
|
|
| ## 📚 Injected Knowledge Stack (Fine-Tuning Data) |
| The model underwent continuous pre-training (*Continual Pre-training*) on **279,049 curated data segments**. To ensure maximum architectural precision and prevent technical overlapping, the data was strictly divided into the following 8 isolated knowledge directories: |
| * **Mobile / Cross-Platform:** Flutter (Latest mobile architecture paradigms and state management). |
| * **Full-Stack Web Architecture:** React Router v8 (Native *Framework Mode* via Vite, strict server loaders, and server actions routing). |
| * **Backend & Runtime Engine:** NestJS & Node.js (API design, server-side middleware, and backend stream logic). |
| * **Database & Persistence Layers:** Prisma ORM & Drizzle ORM (Database schema modeling, safe SQL query builders, and relational mapping). |
| * **Language & System Rigor:** TypeScript (Strict typing guidelines to maximize the model's self-debugging and compiler compliance). |
| * **Design & UI Systems:** Tailwind CSS & Shadcn UI / Radix Primitives (Atomic utility styling directly embedded in JSX/TSX components). |
|
|
| ## 📊 Training Logs & Learning Curve |
| The training ran smoothly over 250 hardware-optimized steps. The training logs demonstrate a clear late convergence phenomenon (the "Eureka" moment) around step 140, where the model successfully bridged mathematical connections and synergies across the 8 isolated documentation stacks. |
|
|
| * **Step 10 (Start):** Loss = `3.643480` (Initial ingestion of complex multi-framework syntaxes) |
| * **Step 50 (Progress):** Loss = `2.829789` |
| * **Step 140 (Logical drop):** Loss = `2.788493` (Synergy and cross-stack architectural understanding) |
| * **Step 250 (Final score):** Loss = `2.508277` (Weight stabilization and general full-stack balance) |
|
|
| ## 💻 Local Execution Guide (Target: GTX 1050 4GB VRAM) |
| For the **[GGUF (Q6_K)](https://huggingface.co/iBotIA/Qwen3.8-4B-Empero-AI-FullStack-GGUF)** build of this model deployed inside **Unsloth Desktop**, **LM Studio**, or **OpenCode**, scrupulously apply these 3 hardware parameters to safeguard your local system resources: |
|
|
| 1. **GPU Offload:** Set the threshold strictly to **25 layers**. This loads ~2.5 GB of the core layers into your **NVIDIA GTX 1050** hardware VRAM, leaving your native Intel/system GPU enough breathing room to manage Windows without lag. The remaining model weight will safely overflow into your 16GB of system RAM. |
| 2. **Sampling Settings:** Set your temperature strictly to **`0.6`** (the standard recommended setting for Empero-AI distillations) along with `top_p=0.95` and `top_k=20`. Do not force greedy decoding (temperature=0), as reasoning models in this class will fall into repetition loops. |
| 3. **Context Window:** Set the context length between **`16384` and `32768` tokens**. This large memory window enables autonomous coding agents to evaluate and rewrite multiple code repository files simultaneously. |
|
|
| ## 🛠️ Automated Software Engineering with OpenCode |
| To deploy this model as a fully autonomous software engineer capable of inspecting, writing, and debugging local app repositories, initialize Unsloth's OpenAI-compatible backend server and hook it to your terminal-based agent: |
|
|
| ```bash |
| unsloth start opencode --context-length 32000 |
| ``` |
|
|
| ## Support / Donate |
|
|
| If this model helped you, consider supporting the project: |
|
|
| - **BTC**: `18cBC5sFjtctw121ULTkxTbTZPurginJBs` |
| - **LTC**: `ltc1q3jrcwrx66xpz4k92p08u8c5v8zwywk3dqpzdkv` |
| - **USDT**: `TGKVpbbznmvEusKbuZZj4WSK6XxtHcG6FE` (TRX chain) |
| - **USDT**: `0x1059cb5a1F8467e5b56a9bdf082cE86FFB002D15` (POL chain) |
| - **USDT**: `0x18b2AA731daeFD47DFFa278f3F856eAF80376fd6` (ETH chain) |
| - **USDT**: `0x3bEcddC7c49bDba5503eB1677628b4519439884c` (BNB chain) |
|
|
| ## Provenance & Licensing |
| The model weights are released under the open and permissive **Apache-2.0** license, inherited from the base Qwen3.5-4B architecture. You are fully free to use, modify, alter, or integrate this model for private enterprise, archival, or commercial software deployment pipelines. |
|
|
