Overview

Jan-v3-4B-base-instruct is a 4 billion parameter instruction-tuned model developed by janhq, built upon the core architecture of Qwen/Qwen3-4B-Instruct-2507. This model was created through post-training distillation from a larger teacher model, allowing it to transfer advanced capabilities while maintaining strong general-purpose performance on standard benchmarks. It is designed to be a compact and easily fine-tunable base model, aiming to reduce the typical compromises between model size and capability.

Key Capabilities

• Compact and Efficient: A 4B parameter model (3.6B non-embedding) with 36 layers, making it suitable for resource-constrained environments.
• Extended Context Length: Features a notable native context length of 262,144 tokens.
• Improved Instruction Following: Offers enhanced instruction adherence out-of-the-box.
• Strong Fine-tuning Base: Provides a robust starting point for downstream fine-tuning tasks.
• Lightweight Coding Assistance: Effective for basic coding support.

Intended Use Cases

• Serving as a superior small base model for various downstream applications.
• Applications requiring improved instruction following capabilities.
• Projects needing a strong foundation for further fine-tuning.
• Use cases benefiting from lightweight coding assistance.