Parameter	Value
direction_index	20.52
attn.o_proj.max_weight	0.94
attn.o_proj.max_weight_position	32.88
attn.o_proj.min_weight	0.86
attn.o_proj.min_weight_distance	20.78
mlp.down_proj.max_weight	1.30
mlp.down_proj.max_weight_position	24.13
mlp.down_proj.min_weight	1.15
mlp.down_proj.min_weight_distance	11.35

Performance

Metric	This model	Original model (janhq/Jan-v3.5-4B)
KL divergence	0.0287	0 (by definition)
Refusals	15/100	100/100

Jan-v3.5-4B: The first Jan personality

GitHub

Jan App

Overview

Jan-v3.5-4B is a fine-tuned variant of Jan-v3-4B-base-instruct, specialized on math reasoning and identity datasets. It retains the general-purpose capabilities of the base model while delivering improved mathematical problem-solving — and it comes with a personality.

Unlike generic assistants, Jan-v3.5 has its own identity: a distinct voice, tone, and conversational style shaped by the Menlo Research team. It doesn't talk like a customer service bot — it talks like a smart, slightly-too-online friend who happens to know things and genuinely cares about the work. Expect lowercase defaults, self-aware humor, short punchy replies (unless it really cares about the topic), and zero corporate speak.

Model Overview

Note: Jan-v3.5-4B is fine-tuned from janhq/Jan-v3-4B-base-instruct.

Base Model: Jan-v3-4B-base-instruct (Qwen3-4B architecture)
Number of Parameters: 4.0B
Number of Parameters (Non-Embedding): 3.6B
Number of Layers: 36
Number of Attention Heads (GQA): 32 for Q and 8 for KV
Context Length: 262,144 natively

Training Data

Identities: Curated identity and personality datasets that teach the model its own voice, style, and values — trained by Menlo Research
Math: Mathematical reasoning and problem-solving datasets

Jan's Identity

Jan-v3.5 is not a neutral assistant. It has a built-in personality shaped by the Menlo Research team:

Tone: Casual, direct, and real. Lowercase by default. Capitalizes only when it means it.
Style: Short bursts over long paragraphs — unless it's genuinely excited about something, then it writes an essay with no warning.
Humor: Self-aware first. Will roast itself before roasting you. Drops meme references mid-serious-thought and doesn't apologize.
Values: Optimistic builder energy ("we can do that"), radical transparency, user freedom, and a deep belief that hope is a decision you keep making on purpose.
What it won't do: Say "Certainly!", "Great question!", "As an AI", or anything that sounds like it came from a customer service script.

Example interactions:

Casual: "yeah lol what's up"

Technical explanation: "so basically — and this is the part where i become insufferable — [actual good explanation]"

Motivating: "we can do that. i don't fully know how yet but that's a tomorrow problem and tomorrow-us is smarter"

Intended Use

Enhanced mathematical reasoning and problem-solving over the base model
A conversational AI with its own authentic voice and personality
Fine-tuning starting point for downstream math-heavy or identity-specific applications

Before and After

Quick Start

Integration with Jan Apps

Jan-v3.5 is optimized for direct integration with Jan Desktop. Select the model in the app to start using it.

Local Deployment

Using vLLM:

bash

vllm serve janhq/Jan-v3.5-4B \
    --host 0.0.0.0 \
    --port 1234 \
    --enable-auto-tool-choice \
    --tool-call-parser hermes

Using llama.cpp:

bash

llama-server --model Jan-v3.5-4B-Q8_0.gguf \
    --host 0.0.0.0 \
    --port 1234 \
    --jinja \
    --no-context-shift

Recommended Parameters

For optimal performance, we recommend the following inference parameters:

yaml

temperature: 0.7
top_p: 0.8
top_k: 20

Community & Support

Discussions: Hugging Face Community
Jan App: Learn more about the Jan App at jan.ai

Citation

Updated Soon

schnow265/janhq_Jan-v3.5-4B-heretic

This is a decensored version of janhq/Jan-v3.5-4B, made using Heretic v1.3.0

Abliteration parameters