Jan-v1 Appears Out of Nowhere: Measured Accuracy Surpasses Perplexity Pro, the Strongest 4B Local AI Model on Earth?

The AI field is in turmoil again! The newly launched Jan-v1 model has slightly surpassed the well-known Perplexity Pro with an amazing accuracy of 91.1% in the SimpleQA question-answering evaluation. More importantly, all this powerful performance can be run locally on your personal computer without an internet connection. This article will provide an in-depth analysis of Jan-v1’s technical details, evaluation performance, and a complete installation and operation guide.

The New King of Local AI? Jan-v1 Makes its Debut

In today’s rapidly developing artificial intelligence technology, we are accustomed to equating powerful AI models with cloud servers. But what if I told you that there is a model that can bring top-notch question-answering and search capabilities directly to your personal computer, while maintaining data privacy and the freedom of offline operation? Would you be excited?

This is exactly what Jan-v1 aims to achieve. Recently, a 4B parameter model named Jan-v1 has attracted widespread attention. It is not only a tool fine-tuned for web search and reasoning tasks, but it has also demonstrated amazing strength in public benchmark tests, enough to challenge industry benchmarks.

SimpleQA Evaluation Reveals the Truth: How Does Jan-v1 Dominate the Leaderboard?

Comprehensive accuracy on question answering

Note: Jan-V1, Jan-nano variants and Lucy models were benchmarked with Serper MCP.

Seeing is believing, and the data speaks for itself. In the SimpleQA benchmark test, which measures the comprehensive accuracy of AI question answering, Jan-v1 has delivered an impressive report card.

As you can clearly see from the evaluation chart above, Jan-v1 is at the top of the list with an accuracy of 91.1%. This result is not only amazing, but it also slightly surpasses the well-regarded online AI search engine Perplexity Pro (90.6%).

What does this mean? It means that users can now get question-answering accuracy comparable to or even exceeding top-tier online services in a completely offline local environment. This is undoubtedly a great boon for professionals who value data privacy or need to work in an environment without an internet connection.

It is also worth noting that Jan-v1 also significantly outperforms other well-known models in this evaluation, such as gpt-oss-20B (86.3%), ChatGPT-4.5 (62.5%), and multiple versions of the Gemini and Claude series, fully demonstrating its outstanding performance in the field of question answering.

Unveiling the Powerful Core: Empowered by Qwen3-4B-Thinking

Jan-v1’s power is no accident. Its core is based on the latest Qwen3-4B-Thinking model launched by Alibaba’s Tongyi Qianwen team. This is a model with 4 billion parameters that strikes an excellent balance between being lightweight and high-performance.

The Jan team has built on this foundation and conducted deep fine-tuning for reasoning and tool use. It’s like equipping a smart student with the best tools to solve complex problems more effectively.

In addition, Jan-v1 also supports a context length of up to 256k. To use a simple analogy, this is equivalent to allowing the model to remember and understand the content of a thick book at once, which provides a huge advantage when processing long reports, analyzing complex code, or conducting in-depth conversations.

Experience it Now! How to Run Jan-v1 on Your Computer?

Want to experience the power of Jan-v1 firsthand? The process is simpler than you think. The development team provides a very friendly integrated environment that allows users to get started easily.

Main method of operation: Through the Jan desktop application

1. Download the Jan application: Go to the Jan official website to download and install the application for your operating system.
2. Enable the web search function (key step): To achieve the best question-answering results, you need to enable the search function so that the model can obtain real-time information.
   • Go to Settings → Experimental Features and turn it On.
   • Then go to Settings → MCP Servers and enable the MCP related to search (e.g., Serper).

For advanced users:

If you are used to using other toolchains, Jan-v1 also supports running in llama.cpp and vLLM environments. You can download the model files from Hugging Face:

• Base model: janhq/Jan-v1-4B
• GGUF format (recommended for most local users): janhq/Jan-v1-4B-GGUF

Pro Tips: Optimizing Model Parameter Settings

To unleash the full potential of Jan-v1, the official recommendation is to use the following parameter settings. You can find these setting options on the right side of the conversation window in the Jan application.

• Temperature: 0.6
  • This value controls the creativity of the output. 0.6 can provide more contextually relevant answers while maintaining accuracy, avoiding answers that are too rigid or far-fetched.
• Top P: 0.95
  • Similar to temperature, this parameter is used to filter out words with lower probabilities, making the answers more focused.
• Top K: 20
  • When generating each word, the selection range is narrowed down to the 20 words with the highest probability, which helps to improve the coherence of the answer.
• Max Tokens: 2048
  • Sets the maximum length of a single answer from the model. 2048 is more than enough for most question-answering scenarios.

Conclusion: A New Era for Local AI

The emergence of Jan-v1 is not just the release of another new model. It symbolizes an important trend: top-tier AI performance is gradually moving from the cloud to the endpoint, from centralized to personalized.

It proves that even at a relatively lightweight parameter level like 4B, it is still possible to achieve performance that surpasses large-scale online services through an excellent base model and precise fine-tuning. This opens up new possibilities for the popularization, privacy protection, and customized application of AI. What are you waiting for? Download Jan now and experience the new wave of local AI for yourself!