dcostenco/prism-coder-14b
dcostenco/prism-coder-14b is a 14 billion parameter language model fine-tuned from Qwen3-14B by dcostenco. It specializes in deterministic tool routing for six specific tools and provides comprehensive code assistance, including generation, explanation, and debugging. Optimized for local inference, this model achieves 97% BFCL accuracy on its routing benchmark and is designed for context-aware session management within the Prism AAC system.
Loading preview...
prism-coder:14b Overview
dcostenco/prism-coder-14b is a specialized 14 billion parameter language model, fine-tuned from the Qwen3-14B architecture. Developed by dcostenco, its primary function is to serve as an AAC (Augmentative and Alternative Communication) tool router and code assistant within the Prism AAC system.
Key Capabilities
- Deterministic Tool Routing: Excels at routing across 6 predefined tools, achieving a 97% accuracy on its BFCL routing benchmark.
- Code Assistance: Provides robust capabilities for code generation, explanation, and debugging.
- Context-Aware Session Management: Designed to maintain and utilize session context effectively.
- Efficient Local Inference: The 14B parameter model is optimized to run efficiently on local hardware, fitting within 10GB of RAM.
Training and Deployment
The model was fine-tuned using MLX LoRA on Apple Silicon (M-series) hardware. It is intended for deployment via Ollama, with various GGUF quantizations available, including Q4_K_M, IQ4_XS, and Q3_K_S, offering different size-performance trade-offs. It functions as the primary model in a desktop cascade, preceding larger 32B models or cloud-based solutions like Claude.