Date061125General agent

Jan-v2-VL

An 8B vision-language model that sees the screen and carries out long, many-step tasks in browsers and desktop apps with minimal drift.

2 min read

Overview

Parameters
8B
Base model
Qwen3-VL-8B-Thinking
Modality
Vision-language
Variants
Low, med, high
License
Apache 2.0

49

Long-horizon benchmark · task length (K)

Jan-v2-VL is an 8B vision-language model built on Qwen3-VL-8B-Thinking, designed for long-horizon, multi-step agentic tasks in real software environments — browsers, desktop apps, and beyond. It combines language reasoning with visual perception to execute stable, many-step automation chains with minimal drift.

Try it in Jan

Variants

Jan-v2-VL comes in three variants tuned for different latency/quality tradeoffs:

VariantFocus
Jan-v2-VL-lowEfficiency-oriented, lower latency
Jan-v2-VL-medBalanced latency and quality
Jan-v2-VL-highDeeper reasoning, higher think time

Capabilities

  • Agentic automation & UI control: Follow complex multi-step instructions in browsers and desktop apps
  • Screenshot grounding: Perceive and act on visual state of the screen
  • Tool calling: Native support for BrowserMCP and similar integrations
  • Error recovery: Maintain intermediate state and recover from minor execution errors
  • Long-horizon execution: Stable performance across many-step automation chains

Performance

Long-Horizon Execution

Jan-v2-VL dramatically outperforms competing VLMs on the long-horizon benchmark, completing tasks up to 49 steps in a single turn — nearly 10x more than the next best model.

BenchmarkJan-v2-VLQwen3-VL-8B-ThinkingQwen2.5-VL-7B-InstructGemma3-it-12BLlama3.1-nemotronGLM-4.1V-9B-Thinking
Long-horizon benchmark (Task length (K) executed in a single turn)4952211
Featuring results from the most popular vision-language models (VLMs).

All three Jan-v2-VL variants significantly exceed the Qwen3-VL-8B-Thinking baseline (score: 5):

BenchmarkJan-v2-VL-highJan-v2-VL-medJan-v2-VL-lowQwen3-VL-8B-Thinking (baseline)
Long-horizon execution (Task length (K) executed in a single turn)4937295

Text-Only Benchmarks

Jan-v2-VL maintains strong performance on standard language benchmarks with no degradation from the base model:

BenchmarkQwen3-VL-8B-Thinking (baseline)Jan-v2-VL-lowJan-v2-VL-medJan-v2-VL-high
MMLU74.3874.3674.3874.43
SimpleQA8788.387.187.1
IF Bench – Prompt Acc.0.34010.32650.32310.3639
IF Bench – Instr. Acc.0.37010.36120.35520.3851
AIME250.800.760.670.87
EQ Bench84.484.184.684.1
Creative Writing79.175.0773.074.0
SWE Bench (✓ / 300)101199
Maintaining strong performance across standard language benchmarks.

Multimodal Benchmarks

Core vision capabilities are fully preserved across all variants:

BenchmarkQwen3-VL-8B-Thinking (baseline)Jan-v2-VL-lowJan-v2-VL-medJan-v2-VL-high
RealVQA52.8151.551.751.2
CountBench88.7988.7988.1888.79
OCR bench v2 (EN)0.5820.5830.5830.585
OCR bench v2 (CN)0.5950.5950.5880.609
Ensuring no degradation in core vision capabilities.

Requirements

  • Memory:
    • Minimum: 16GB RAM (with Q4 quantization)
    • Recommended: 24GB RAM (with Q8 quantization)
  • Hardware: GPU recommended for real-time agentic use
  • API Support: OpenAI-compatible at localhost:1337

Using Jan-v2-VL

Quick Start

  1. Download Jan Desktop
  2. Click Open in Jan above, or open the Hub and search for Jan-v2-VL
  3. Choose Jan-v2-VL-med or another variant that fits your hardware, then click Download
  4. Once downloaded, select the downloaded Jan-v2-VL variant from the model selector in any chat
  5. Start using it for browser automation or visual tasks

Deployment Options

Using vLLM:


vllm serve janhq/Jan-v2-VL-high \
--host 0.0.0.0 \
--port 1234 \
--enable-auto-tool-choice \
--tool-call-parser hermes \
--reasoning-parser qwen3

Using llama.cpp:


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


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

What Jan-v2-VL Does Well

  • UI automation: Navigate and interact with browsers and desktop applications via screenshots
  • Long-horizon tasks: Execute multi-step plans without losing context or drifting
  • Visual grounding: Identify and interact with specific UI elements from screenshots
  • Tool calling: Integrate with MCP servers like BrowserMCP for end-to-end automation

Limitations

  • GPU recommended: Real-time agentic use benefits significantly from GPU acceleration
  • 8B size: More capable than 4B models for visual tasks, but larger resource footprint
  • Vision required: Text-only tasks do not benefit from the vision-language architecture

Downloads

GGUF quantizations

  • Q4_K_M: Good balance of size and quality
  • Q5_K_M: Better quality, slightly larger
  • Q8_0: Highest quality quantization

Talk to us