megabytes/Jan-v3-4B-base-instruct-heretic

TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Feb 10, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

megabytes/Jan-v3-4B-base-instruct-heretic is a 4 billion parameter instruction-tuned language model, derived from janhq/Jan-v3-4B-base-instruct and based on the Qwen3-4B-Instruct-2507 architecture. This model has been decensored using the Heretic v1.2.0 tool, significantly reducing refusals compared to its original counterpart. It features a native context length of 262,144 tokens and is intended as a compact, ownable base for fine-tuning, offering improved instruction following and lightweight coding assistance.

Loading preview...

Model Overview

megabytes/Jan-v3-4B-base-instruct-heretic is a 4 billion parameter instruction-tuned model, built upon the Qwen3-4B-Instruct-2507 architecture. It is a decensored version of janhq/Jan-v3-4B-base-instruct, created using the Heretic v1.2.0 tool. This modification significantly reduces refusal rates, with the model exhibiting 17 refusals out of 100 test cases, compared to 100 refusals for the original model.

Key Capabilities

  • Decensored Output: Provides responses with significantly fewer refusals than its base model.
  • Compact & Efficient: A 4B parameter model (3.6B non-embedding) designed for broad applicability.
  • Extended Context: Features a native context length of 262,144 tokens.
  • Strong Base for Fine-tuning: Offers improved instruction following out-of-the-box, making it an effective starting point for downstream tasks.
  • Coding Assistance: Provides lightweight coding assistance.

Intended Use Cases

  • Fine-tuning: Ideal as a small, adaptable base model for custom fine-tuning efforts.
  • Applications Requiring Reduced Refusals: Suitable for use cases where the original model's refusal rate was a limitation.
  • General-Purpose Instruction Following: Effective for various instruction-based tasks due to its distilled capabilities.
  • Lightweight Coding: Can assist with coding tasks where a smaller model is preferred.