dcostenco/prism-coder-7b

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Apr 28, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

dcostenco/prism-coder-7b is a 7 billion parameter instruction-tuned Qwen2.5-Coder-7B model developed by dcostenco/Synalux AI, specifically fine-tuned for on-device function calling and Augmentative and Alternative Communication (AAC) tasks. It excels at AAC realignment, live tool-call accuracy, multi-tool-call parsing, and irrelevance detection, designed for HIPAA-safe local inference on M-series Mac or RTX 30+ GPUs. This model is optimized for specialized applications like the PrismAAC consumer app and Synalux portal, providing robust performance in its target categories.

Loading preview...

Model Overview

dcostenco/prism-coder-7b is a specialized 7 billion parameter model, fine-tuned from Qwen2.5-Coder-7B-Instruct by dcostenco/Synalux AI. It is designed for on-device function calling and Augmentative and Alternative Communication (AAC) applications, particularly for the Synalux platform. This model prioritizes category-specific performance over general-purpose benchmarks, offering robust capabilities for its intended use cases.

Key Capabilities & Differentiators

  • AAC Realignment: Achieves 97.9% accuracy for Augmentative & Alternative Communication app tasks, crucial for nonverbal users and BCBA caregivers.
  • Live Tool-Call Accuracy: Demonstrates 63.21% live accuracy on BFCL V4 for on-device function-calling agents, with 73.28% Live AST and 79.48% Non-Live AST.
  • Multi-Tool-Call Parsing: Capable of parsing multiple tool calls within a single user turn, with 67.43% Live Multiple AST and 82.50% Non-Live Multiple AST.
  • Irrelevance Detection: Scores 69.89% in knowing when not to call a tool, enhancing agent reliability.
  • Emergency QA: Maintains 100% accuracy (13/13 cases) for parsing emergency questions, vital for nonverbal users in critical situations.
  • Local Inference: Designed for HIPAA-safe, on-device operation, eliminating cloud round-trips and running efficiently on M-series Macs or RTX 30+ GPUs.

Ideal Use Cases

This model is specifically tailored for:

  • PrismAAC Consumer App: Serving as the on-device function-calling backbone for AAC applications, handling caregiver instructions, text correction for motor-impaired users, emergency QA, translation across 12 languages, and open-ended Q&A.
  • Synalux Portal: Enabling tier-aware routing for standard queries locally, significantly reducing cloud costs for high-scale user bases by offloading simpler requests from more expensive cloud-based models.