|
|
--- |
|
|
license: apache-2.0 |
|
|
base_model: |
|
|
- Qwen/Qwen2.5-Coder-7B-Instruct |
|
|
- deepseek-ai/DeepSeek-R1-Distill-Qwen-7B |
|
|
tags: |
|
|
- merge |
|
|
- mergekit |
|
|
- ties |
|
|
- coding |
|
|
- reasoning |
|
|
- aesir-unlimited |
|
|
- web3 |
|
|
--- |
|
|
|
|
|
# ⚡ AESIR-Coder-7B |
|
|
|
|
|
**AESIR-Coder-7B** is a high-performance, reasoning-dense language model developed by **ÆSIR Unlimited**. By merging the world-class syntax precision of **Qwen2.5-Coder** with the rigorous Chain-of-Thought (CoT) logic of **DeepSeek-R1**, we have engineered a "pocket-sized architect" capable of handling complex software engineering and Web3 tasks on consumer-grade hardware. |
|
|
|
|
|
## 🛠️ Architecture & Methodology |
|
|
|
|
|
This model was engineered using the **TIES (Trimming, Electing, and Merging)** method via `mergekit`. This approach allows the model to resolve weight conflicts between its two parent brains—ensuring the reasoning logic of DeepSeek doesn't "break" the specific coding syntax of Qwen. |
|
|
|
|
|
- **Base Engine:** Qwen2.5-Coder-7B-Instruct (Optimized for 90+ programming languages) |
|
|
- **Reasoning Layer:** DeepSeek-R1-Distill-Qwen-7B (Trained on massive reasoning traces) |
|
|
- **Merge Method:** TIES |
|
|
- **Hardware Profile:** Optimized for 8GB-16GB RAM environments (Local First) |
|
|
|
|
|
## 🚀 Key Agentic Capabilities |
|
|
|
|
|
1. **Chain-of-Thought Auditing:** Unlike standard models that just write code, AESIR-Coder "thinks" through the logic. It is ideal for identifying logic flaws in smart contracts and complex Python agentic systems. |
|
|
2. **Web3 Native:** Deep knowledge of Solidity, Vyper, and the TON/Solana ecosystems, combined with the ability to reason about decentralized state-machines. |
|
|
3. **Structured Intelligence:** Highly stable at generating JSON schemas and YAML configurations required for the **ÆSIR Protocol** agentic handshakes. |
|
|
|
|
|
## 💻 Local Execution (ÆSIR Unlimited Standards) |
|
|
|
|
|
To run this model on your local machine with limited RAM, we recommend using **LM Studio** or **Ollama** with a **Q4_K_M GGUF** quantization. |
|
|
|
|
|
| RAM Availability | Recommended Quantization | |
|
|
| :--- | :--- | |
|
|
| **8GB RAM** | Q4_K_M (Fast, High Quality) | |
|
|
| **12GB+ RAM** | Q6_K or Q8_0 (Near-Lossless) | |
|
|
|
|
|
--- |
|
|
*Developed by **Bugg-Moran** as part of the **ÆSIR Unlimited** mission to bridge the gap between AI and Decentralized Infrastructure.* |
|
|
|
