KBBridge/KBBridge-v3

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 27, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

KBBridge-v3 is a 27 billion parameter language model developed by KBBridge, fine-tuned from Qwen3.8-27B, specifically for generating and understanding GeneXus programming code in the native .gxSource format. It excels at producing syntactically valid GeneXus objects and explaining existing code, offering a specialized solution for GeneXus developers. The model features a 32768-token context length and can run on local hardware, ensuring code privacy.

Loading preview...

KBBridge-v3: Specialized GeneXus Code Assistant

KBBridge-v3 is a 27 billion parameter model, fine-tuned from Qwen3.8-27B, designed to assist GeneXus developers. Its primary specialization is generating and interpreting GeneXus programming code in the .gxSource format, a task where general-purpose LLMs typically struggle.

Key Capabilities & Differentiators

  • GeneXus Code Generation: Achieves an 84.8% parseRate for .gxSource syntax, significantly outperforming frontier models that produce syntactically invalid output without extensive prompting.
  • Code Explanation: When reasoning is enabled, it nearly halves the rate of fabricated claims when explaining existing code (from 15.4% to 8.7%).
  • Privacy-Focused: Operates locally, ensuring GeneXus Knowledge Base code is never sent to external APIs. Training data audits showed no memorization of customer code, even with 1000 repetitions of synthetic canaries.
  • Optimized for GeneXus: Internalizes GeneXus documentation, making external documentation injection counterproductive.
  • Flexible Reasoning: Users can toggle reasoning on/off depending on the task; off for code generation, on for code explanation.
  • Quantization Options: 4-bit quantized builds (e.g., FP8, GGUF) offer comparable output quality with reduced memory footprint.

Important Usage Notes

  • Reasoning Setting: Turn reasoning OFF for .gxSource generation to avoid significant degradation in parse rate. Turn it ON for explaining existing code.
  • Explicit Format Request: Always include "in .gxSource format" in prompts for consistent GeneXus output.
  • Adequate Context: Use max_tokens ≥ 4096, as .gxSource objects consume approximately 340 tokens per KB of source.

Limitations

  • No KB Knowledge: The model does not know the contents of specific Knowledge Bases; it will invent plausible attribute names if not provided context.
  • Not General-Purpose: Worse than its base model at general tasks.
  • Runaway Generation: Can exhibit degenerate repetition for very large objects (over ~10 KB).

KBBridge-v3 offers a robust, specialized solution for GeneXus development, prioritizing accuracy in a niche domain and data privacy.

Finetunes
3 models