The Jan team has announced the release of Jan-v3-4B-base-instruct, a model with 4 billion parameters developed through continual pre-training and reinforcement learning (RL).

Key Features

  • Objective: Improve performance in common tasks while maintaining the general capabilities of the model.
  • Usage: Excellent starting point for fine-tuning and for improving lightweight assistance in math and coding.
  • How to run it: Available via Jan Desktop (downloadable from the official website) and on Hugging Face.
  • Recommended parameters:
    • Temperature: 0.7
    • top_p: 0.8
    • top_k: 20

Upcoming developments

The Jan team has announced that the following are coming soon:

  • Jan-Code: a finetuned version of Jan-v3-4B-base-instruct focused on coding.
  • Jan-v3-Search-4B: an update of Jan-nano based on Jan-v3-4B-base-instruct.
  • A family of 30B parameter Jan-v3 models.

For those evaluating on-premise deployments, AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate trade-offs.