janhq/Jan-v1-edge
Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:2BQuant:BF16Ctx Length:32kPublished:Aug 27, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

Jan-v1-edge is a 2 billion parameter agentic language model developed by janhq, distilled from the larger Jan-v1 model. Optimized for fast, reliable on-device execution, it features a 40960 token context length and is designed for resource-constrained environments. This model excels in reasoning and problem-solving, making it suitable for web search and interactive workloads.

Loading preview...

Overview

Jan-v1-edge is a lightweight, agentic language model from janhq, designed for efficient on-device execution. Distilled from the larger Jan-v1 model, it retains strong reasoning and problem-solving capabilities within a smaller 2 billion parameter footprint, making it ideal for resource-constrained environments.

Key Capabilities & Features

  • Distilled Architecture: Developed through a two-phase post-training process involving Supervised Fine-Tuning (SFT) from Jan-v1 and Reinforcement Learning with Verifiable Rewards (RLVR).
  • High Efficiency: Achieves 83% accuracy on SimpleQA, demonstrating robust performance despite its small size.
  • Agentic Design: Built for interactive workloads and web search applications.
  • Optimized for Edge: Specifically engineered for fast, reliable execution on local devices.
  • Context Length: Supports a substantial 40960 token context window.

Performance Highlights

While showing a slight degradation in instruction-following and creative writing compared to Qwen 3 1.7B Thinking, Jan-v1-edge remains comparable or superior in EQBench and recency-based Question Answering.

Integration

Jan-v1-edge is optimized for direct integration with the Jan App and supports local deployment via vLLM and llama.cpp.