Date080825General agent

Jan-v1

A 4B reasoning model built on Qwen3-4B-thinking that scores 91.1% on SimpleQA, made for question answering and problem-solving.

2 min read

Overview

Parameters
4B
Base model
Qwen3-4B-Thinking
Built for
Reasoning and problem-solving
License
Apache 2.0

91.1%

SimpleQA accuracy

Jan-v1 is a 4B parameter model based on Qwen3-4B-thinking, designed for reasoning and problem-solving tasks. The model achieves 91.1% accuracy on SimpleQA through model scaling and fine-tuning approaches.

Try it in Jan

Performance

SimpleQA Benchmark

Jan-v1 demonstrates strong factual question-answering capabilities:

BenchmarkJan-v1Perplexity ProQwen3-4B-2507gpt-oss-20BJan-nano-128kJan-nanoJan-nano (YaRN)Lucy (YaRN)DeepSeek-V3ChatGPT-4.5BaselineGemini-2.5-ProClaude-3.7-Sonneto3Grok-3o1
SimpleQA (Question answering accuracy)91.1%90.6%86.5%86.3%83.2%80.7%79.7%78.3%78.2%62.5%59.2%52.9%50.0%49.4%44.6%42.6%
Jan-v1, the Jan-nano variants and Lucy are benchmarked using Serper MCP.

At 91.1% accuracy, Jan-v1 outperforms several larger models on SimpleQA, including Perplexity's 70B model. This performance represents effective scaling and fine-tuning for a 4B parameter model.

Chat and Creativity Benchmarks

Jan-v1 has been evaluated on conversational and creative tasks:

BenchmarkJan-v1Qwen3-4B-ThinkingGPT-OSS-20B (High)GPT-OSS-20B (Low)
EQBench83.6182.6178.3578.35
CreativeWriting72.0865.7430.2326.38
IFBench39.148.0660.054.03

These benchmarks (EQBench, CreativeWriting, and IFBench) measure the model's ability to handle conversational nuance, creative expression, and instruction following.

Requirements

  • Memory:
    • Minimum: 8GB RAM (with Q4 quantization)
    • Recommended: 16GB RAM (with Q8 quantization)
  • Hardware: CPU or GPU
  • API Support: OpenAI-compatible at localhost:1337

Using Jan-v1

Quick Start

  1. Download Jan Desktop
  2. Click Open in Jan above, or open the Hub and search for Jan-v1
  3. Choose a variant that fits your hardware, then click Download
  4. Once downloaded, select Jan-v1 from the model selector in any chat
  5. Start chatting - no additional configuration needed

Demo

Jan-v1 Demo

Deployment Options

Using vLLM:


vllm serve janhq/Jan-v1-4B \
--host 0.0.0.0 \
--port 1234 \
--enable-auto-tool-choice \
--tool-call-parser hermes

Using llama.cpp:


llama-server --model jan-v1.gguf \
--host 0.0.0.0 \
--port 1234 \
--jinja \
--no-context-shift


temperature: 0.6
top_p: 0.95
top_k: 20
min_p: 0.0
max_tokens: 2048

What Jan-v1 Does Well

  • Question Answering: 91.1% accuracy on SimpleQA
  • Reasoning Tasks: Built on thinking-optimized base model
  • Tool Calling: Supports function calling through hermes parser
  • Instruction Following: Reliable response to user instructions

Limitations

  • Model Size: 4B parameters limits complex reasoning compared to larger models
  • Specialized Tasks: Optimized for Q&A and reasoning, not specialized domains
  • Context Window: Standard context limitations apply

Downloads

GGUF quantizations

  • Q4_K_M: 2.5 GB - Good balance of size and quality
  • Q5_K_M: 2.89 GB - Better quality, slightly larger
  • Q6_K: 3.31 GB - Near-full quality
  • Q8_0: 4.28 GB - Highest quality quantization

Talk to us