intrect/VELA

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

VELA (Vector-Encoded Learning Agent) by intrect is a 7.6 billion parameter language model, fine-tuned from Qwen/Qwen2.5-7B-Instruct, specifically optimized for Korean stock market news analysis and investment research. It excels at analyzing news impact on 2,135 KOSPI/KOSDAQ stocks, interpreting securities reports, and generating structured investment analysis using a Reasoning Trace. Trained with over 58K SFT samples and 26K DPO pairs, VELA provides accurate and structured analysis within the Korean financial domain.

Loading preview...

VELA: Korean Stock Market AI Analyst

VELA is a 7.6 billion parameter language model developed by intrect, specialized in analyzing the Korean stock market. Built upon the Qwen/Qwen2.5-7B-Instruct base model, VELA is fine-tuned with over 58,000 Supervised Fine-Tuning (SFT) samples and 26,000 Direct Preference Optimization (DPO) pairs, focusing on financial domain data.

Key Capabilities

  • News Impact Analysis: Systematically infers the impact of stock-related news on prices.
  • Reasoning Trace: Generates transparent, step-by-step analysis in Markdown format, showing the thought process.
  • Securities Report Interpretation: Extracts key insights and investment implications from analyst reports.
  • Structured Investment Reports: Produces comprehensive reports across seven sections, including Executive Summary, Market Trends, and Investment Opinion.
  • Tool Calling: Integrates with external tools like Search, Price, and Investor for real-time data access (requires agent framework).

Performance & Optimization

VELA demonstrates strong performance in Korean financial contexts, showing improved scores in financial-related subjects on KMMLU benchmarks compared to its base model. It maintains base model capabilities without catastrophic forgetting. The model is available in BF16, GGUF (Q4_K_M), and MLX 4-bit formats, with MLX offering 3.2x faster inference and 73% reduced model size on Apple Silicon. Extensive DPO training has significantly reduced Chinese/English language leaks and improved Reasoning Trace format adherence to ~98%.

Ideal Use Cases

  • Automated analysis of Korean stock market news.
  • Generating structured investment research reports.
  • Interpreting complex financial documents and analyst reports.
  • Developing AI agents for financial decision support in the Korean market.

Finetunes
1 models