HeshamXOR/cubex-flash

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 24, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Cubex-Flash is a 12-billion parameter LoRA fine-tune of Google's Gemma-4-12B, specifically optimized for agentic tool use, reasoning, and instruction-following with a 32768-token context length. Developed by HeshamXOR, it was trained on a unique 4-pillar dataset mixture to enhance capabilities in agentic coding, step-by-step reasoning, and function calling. This model demonstrates statistically significant improvements in tool-calling and constraint-following compared to its base model, making it suitable for applications requiring precise agentic behavior.

Loading preview...

Cubex-Flash Overview

Cubex-Flash is a 12-billion parameter LoRA fine-tune of google/gemma-4-12B, developed by HeshamXOR. It is specifically designed to excel in agentic tool use, reasoning, and instruction-following, leveraging a 32768-token context length. This model was trained on a unique 4-pillar dataset mixture, including agentic coding reasoning, test-time-scaling reasoning, agentic tool use, and constraint alignment, totaling nearly 9,000 examples.

Key Capabilities

  • Enhanced Agentic Tool Use: Demonstrates statistically significant improvements in function calling, achieving 99.0% name match and 81.0% strict match on the xLAM tool-calling benchmark.
  • Improved Constraint Following: Shows a statistically significant gain in adhering to instructions and constraints, scoring 55.0% on a custom 20-prompt suite.
  • Reasoning Potential: Exhibits a promising, though not yet statistically significant, gain in GSM8K reasoning (63.0% vs. 51.0% for base).
  • Faster Inference: Benchmarks indicate significantly faster runtime compared to the base model, completing the same suite in 115 minutes versus 344 minutes for the base, due to more concise and token-efficient completions.

Good For

  • Agentic Applications: Ideal for developing AI agents that require robust tool-calling and function execution capabilities.
  • Instruction-Following Tasks: Suitable for scenarios where precise adherence to complex instructions and constraints is critical.
  • Reasoning-Intensive Workloads: Can be considered for tasks demanding step-by-step reasoning, with potential for further improvement.

Note: This model is released under a CC-BY-NC-4.0 (non-commercial) license due to its training data. Users requiring commercial terms should review the licensing of individual datasets or consider retraining without the restrictive components.

Finetunes
1 models