Date200326General agent

Jan-v3.5-4B

The first Jan personality: a 4B model fine-tuned for math reasoning, with a conversational voice of its own.

1 min read

Overview

Parameters
4.0B (3.6B non-embedding)
Fine-tuned from
Jan-v3-4B-base-instruct
Training focus
Math reasoning, personality
License
Apache 2.0

262K

Context length · 262,144 tokens

Jan-v3.5-4B is the first Jan personality: a 4-billion parameter fine-tuned variant of Jan-v3-4B-base-instruct, specialized in mathematical reasoning while incorporating a distinct conversational personality developed by Menlo Research.

This model enhances the base Qwen3-4B architecture with improved math problem-solving capabilities and a deliberately crafted identity. Rather than functioning as a neutral assistant, Jan-v3.5-4B uses casual, direct language with lowercase-by-default styling and maintains an authentic conversational tone.

Try it in Jan

Key Features

  • Base architecture: Qwen3-4B with 36 layers and 4.0B parameters (3.6B non-embedding)
  • Context length: 262,144 tokens natively supported
  • Training focus: math reasoning datasets and curated identity/personality datasets
  • Distinctive voice: self-aware humor, direct communication style, avoids corporate phrasing
  • Attention mechanism: 32 query heads with 8 key-value heads (GQA)

Jan-v3.5-4B before/after comparison


temperature: 0.7
top_p: 0.8
top_k: 20

Usage

Jan-v3.5-4B supports deployment via:

  • Transformers library with pipeline or direct model loading
  • vLLM server with tool-calling capabilities
  • llama.cpp with quantized variants
  • Jan Desktop application integration

Limitations

  • Optimized specifically for math-heavy tasks over general purpose
  • Personality training may not align with all use cases
  • Intended as a starting point for downstream fine-tuning

Downloads


Talk to us