---
license: apache-2.0
language:
- en
base_model:
- Qwen/Qwen3-VL-30B-A3B-Thinking
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- agent
---
# Jan-v2-VL: Multimodal Agent for Long-Horizon Tasks

[![GitHub](https://img.shields.io/badge/GitHub-Repository-blue?logo=github)](https://github.com/janhq/jan) 
[![License](https://img.shields.io/badge/License-Apache%202.0-yellow)](https://opensource.org/licenses/Apache-2.0)
[![Jan App](https://img.shields.io/badge/Powered%20by-Jan%20App-purple?style=flat&logo=android)](https://jan.ai/) 


![image](https://cdn-uploads.huggingface.co/production/uploads/655e3b59d5c0d3db5359ca3c/NyZR7QqxeifhCXKLQUapu.png)

## Overview

**Jan-v2-VL-max** extends the Jan-v2-VL family to a **30B-parameter** vision–language model focused on **long-horizon execution**. This release scales model capacity and applies **LoRA-based RLVR** to improve stability over many steps with **low error accumulation**. For evaluation, we continue to use **[The Illusion of Diminishing Returns: Measuring Long-Horizon Execution in LLMs](https://arxiv.org/pdf/2509.09677)**, which emphasizes execution length rather than knowledge recall.
### Intended Use
Tasks where the plan and/or knowledge can be provided up front, and success hinges on stable, many-step execution with minimal drift:

* **Agentic automation & UI control:** Stepwise operation in browsers/desktop apps with screenshot grounding and tool calls via **[Jan Browser MCP](https://chromewebstore.google.com/detail/jan-browser-mcp/mkciifcjehgnpaigoiaakdgabbpfppal?pli=1)**.


## Model Performance

Evaluated under FP8 inference, **Jan-v2-VL-max** vs. **[Qwen3-VL-30B-A3B-Thinking](https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Thinking)** shows **no regressions** and **small gains** on several tasks, with the largest improvements in **long-horizon execution**. Our FP8 build maintains accuracy while reducing memory footprint and latency.

![image](https://cdn-uploads.huggingface.co/production/uploads/655e3b59d5c0d3db5359ca3c/pRHKkb-YRyQmimKiQuO4w.png)

## Local Deployment

### Jan Web

Hosted on **Jan Web** — use the model directly at **[chat.jan.ai](https://chat.jan.ai/)** 

![image/gif](demo.gif)

### Local Deployment

**Using vLLM:**
We recommend **vLLM** for serving and inference. All reported results were run with **vLLM 0.12.0**.
For **FP8** deployment, we used **llm-compressor** built from source. Please pin transformers==4.57.1 for compatibility.

```bash
# Exact versions used in our evals
pip install vllm==0.12.0
pip install transformers==4.57.1
pip install "git+https://github.com/vllm-project/llm-compressor.git@1abfd9eb34a2941e82f47cbd595f1aab90280c80"
```

```bash
vllm serve Menlo/Jan-v2-VL-max-FP8 \
    --host 0.0.0.0 \
    --port 1234 \
    -dp 1 \
    --enable-auto-tool-choice \
    --tool-call-parser hermes \
    --reasoning-parser deepseek_r1 
    
```

### Recommended Parameters
For optimal performance in agentic and general tasks, we recommend the following inference parameters:
```yaml
temperature: 1.0
top_p: 0.95
top_k: 20
repetition_penalty: 1.0
presence_penalty: 1.5
```

## 🤝 Community & Support

- **Discussions**: [Hugging Face Community](https://huggingface.co/janhq/Jan-v2-VL-max-FP8/discussions) 
- **Jan App**: Learn more about the Jan App at [jan.ai](https://jan.ai/)

## 📄 Citation
```bibtex
Updated Soon
```