dracko14/Myth
dracko14/Myth is a 1.54 billion parameter Qwen2-based language model created by dracko14, featuring a 32,768 token context length. It is a SLERP-fused model combining the reasoning capabilities of DeepSeek-R1, the coding precision of Qwen-Coder, and the mathematical rigor of Qwen-Math. This compact model is optimized for efficient edge deployment and low-latency applications, delivering performance comparable to larger 3B models.
Loading preview...
MYTH-1.5B: A Fused Reasoning and Coding Model
MYTH-1.5B, developed by dracko14, is a compact 1.54 billion parameter language model built on the Qwen2 architecture with an extensive 32,768 token context length. Its unique strength lies in its creation via Spherical Linear Interpolation (SLERP) fusion, combining three specialized base models: DeepSeek-R1-Distill-Qwen-1.5B for deep reasoning, Qwen2.5-Coder-1.5B-Instruct for coding precision, and the Qwen-Math lineage for mathematical rigor.
Key Capabilities & Performance
This fusion approach results in a balanced capability profile, excelling in:
- Reasoning: Inherits strong chain-of-thought from DeepSeek-R1.
- Coding: Demonstrates robust code understanding and generation, with an estimated HumanEval score of 52.0.
- Mathematics: Achieves strong performance in math word problems (GSM8K 79.0) and competition mathematics (MATH-500 72.0).
- Instruction Following: Benefits from clean alignment inherited from its source models.
Efficiency and Use Cases
MYTH-1.5B is designed for best-in-class efficiency, offering performance comparable to 3B models at half the size. This makes it particularly suitable for:
- Edge deployment on resource-constrained devices.
- Mobile inference through quantization (e.g., GGUF).
- Low-latency applications where speed is critical.
- Cost-effective serving due to its smaller footprint.
The model's fusion method, implemented via a custom Python engine, allows for rapid model merging without retraining, processing tensors independently for low RAM usage. While powerful for its size, users should note its 1.5B parameter scale means it may not compete with larger models on knowledge-heavy tasks, and independent evaluation of its estimated benchmark scores is recommended.