mainbrains/phi-3-mini-4k-instruct

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:4kPublished:Aug 20, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

The Phi-3-Mini-4K-Instruct is a 3.8 billion parameter, instruction-tuned causal language model developed by Microsoft. It is part of the Phi-3 family, trained on high-quality, reasoning-dense datasets, and features a 4K token context length. This model excels in strong reasoning, particularly math and logic, and is optimized for memory/compute-constrained and latency-bound environments.

Loading preview...

Model Overview

The mainbrains/phi-3-mini-4k-instruct is a 3.8 billion parameter, instruction-tuned model from Microsoft's Phi-3 family, specifically the Mini variant with a 4K token context length. It was trained using Phi-3 datasets, which include synthetic data and filtered public web data, emphasizing high-quality and reasoning-dense properties. The model underwent supervised fine-tuning (SFT) and direct preference optimization (DPO) for improved instruction following and safety.

Key Capabilities & Performance

  • Strong Reasoning: Achieves robust performance on benchmarks testing common sense, language understanding, math, code, and logical reasoning, often comparable to models larger than 13 billion parameters.
  • Instruction Following: A June 2024 update significantly improved instruction following, structured output generation (JSON, XML), and multi-turn conversation quality.
  • Efficiency: Designed for memory/compute-constrained environments and latency-bound scenarios, making it suitable for edge hardware.
  • Benchmark Highlights: Scores 70.9 on MMLU, 85.7 on GSM8K CoT, and 57.3 on HumanEval, demonstrating strong capabilities in language, math, and code.

Ideal Use Cases

  • Resource-Constrained Environments: Excellent for applications requiring low latency on edge hardware (e.g., Raspberry Pi 5, Jetson).
  • Reasoning Tasks: Particularly strong in mathematical and logical reasoning.
  • Structured Output: Highly effective for generating structured outputs like JSON and XML.
  • Multi-Agent Pipelines: Can serve as a lightweight reasoning backbone for multi-agent systems, handling tool-calling and structured output generation efficiently.