Jan-v2-VL: Multimodal Agent for Long-Horizon Tasks

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, 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.

Model Performance

Evaluated under FP8 inference, Jan-v2-VL-max vs. 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.

Local Deployment

Jan Web

Hosted on Jan Web — use the model directly at chat.jan.ai

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.

# 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"

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 
    

Using llama.cpp:

llama-server --model Jan-v2-VL-max-Q8_0.gguf \
    --vision-model-path Jan-v2-VL-max-mmproj.gguf \
    --host 0.0.0.0 \
    --port 1234 \
    --jinja \
    --no-context-shift

Recommended Parameters

For optimal performance in agentic and general tasks, we recommend the following inference parameters:

temperature: 1.0
top_p: 0.95
top_k: 20
repetition_penalty: 1.0
presence_penalty: 1.5

🤝 Community & Support

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

📄 Citation

Updated Soon