Phinance-Phi-4-mini-instruct-finance-v0.4-with-reasoning Overview

This model, developed by Joseph G. Flowers, is a compact 3.8 billion parameter instruction-tuned variant of Microsoft's Phi-4-mini-instruct, specialized for financial applications. It integrates financial QA, reasoning chains, and RAG-ready formatting, making it suitable for advanced finance tasks. A key differentiator is its ability to produce structured outputs, follow complex instructions, and chain logical steps, especially when prompted with tags like <thinking>.

The model has undergone extensive fine-tuning, including a recent "Finance Curriculum Reasoning Expansion" phase. This involved training on newly released multilingual finance reasoning datasets, such as Finance Curriculum Edu English and others in Arabic, Uzbek, and a general multilingual set. This training specifically targeted real-world coverage gaps in non-English finance QA, enhancing conceptual reasoning and QA coherence across over 60 languages, and improving robustness to diverse phrasing and financial domains.

Key Capabilities

Financial Reasoning: Excels at multi-step reasoning for investment strategies, reports, and economic topics.
Instruction Following: Provides precise response formatting based on few-shot examples or system messages.
Structured Output: Capable of generating valid JSON for tasks like entity extraction, parsing, and tagging.
RAG-Compatible: Efficiently handles external context prepended in the user field.
Tag-Aware Reasoning: Supports <thinking> tags to guide and display reasoning chains.
Multilingual Finance QA: Offers expanded coverage for curriculum-based financial topics in 60+ languages.

Good for

Developers building financial agents requiring structured, reasoned outputs.
Applications needing precise instruction following in financial contexts.
Use cases involving multi-turn financial dialogues.
Tasks requiring financial analysis and QA in multiple languages, particularly non-English settings.