dcostenco/prism-coder-9b
dcostenco/prism-coder-9b is a QLoRA fine-tuned Qwen3.5-9B model developed by dcostenco, specifically optimized for MCP (Multi-Component Program) tool routing within the Prism Coder system. This model features a hybrid attention architecture combining linear and full attention layers, enabling efficient pattern matching and precise reasoning. It achieves 100% accuracy across multiple categories on the BFCL benchmark, demonstrating high reliability and zero hallucinations for its specialized task. Designed as the default router in the Prism Coder fleet, it offers improved accuracy and a smaller footprint compared to its 14B predecessor.
Loading preview...
Overview
dcostenco/prism-coder-9b is a specialized language model, QLoRA fine-tuned from the Qwen3.5-9B base model. Developed by dcostenco, its primary function is MCP (Multi-Component Program) tool routing within the Prism Coder system. This model is designed to replace previous, larger iterations, offering a more efficient solution with enhanced performance.
Key Capabilities & Features
- Exceptional Tool Routing Accuracy: Achieves a perfect 100% accuracy across 64 shuffled runs on the BFCL benchmark, with zero hallucinations. This significantly surpasses the 90.3% of its 14B predecessor.
- Optimized Footprint: At 5.8 GB, it is 36% smaller than the previous 14B model (9.0 GB) while delivering higher accuracy.
- Hybrid Attention Architecture: Leverages Qwen3.5-9B's unique architecture, combining linear attention layers (Gated DeltaNet) for O(n) inference and pattern matching, with full attention layers for precise retrieval and reasoning.
- Vision Capability: Inherits the Qwen3.5 vision tower, allowing it to accept image inputs. Video is processed as frames through this vision tower.
- Robust Training: Fine-tuned using QLoRA (4-bit base + bf16 adapters) on a corpus of 26K rows, including tool-use, AAC, abstention, and safety data.
Ideal Use Cases
- Default Tool Router: Specifically designed to serve as the default router within the Prism Coder fleet, handling complex tool remapping, parameter normalization, and multi-turn chain parsing.
- Applications Requiring High Accuracy: Suitable for scenarios where precise tool selection and execution are critical, backed by its 100% BFCL benchmark score.
- Resource-Constrained Environments: Its optimized size makes it a strong candidate for deployments where model footprint is a concern, without compromising on performance.