Jan-v2-VL-med-FP8

Jan-v2-VL-med-FP8

Jan-v2-VL-med-FP8 is an FP8-compressed variant built on top of janhq/Jan-v2-VL-med. This edition applies BF16 · FP8 (F8_E4M3) precision formats to significantly reduce memory footprint and improve inference throughput while preserving the structured tool-calling and multimodal agent capabilities of the original 8B architecture. The base Jan-v2-VL-med model from JanHQ is the balanced variant in the 8B-parameter vision-language model family built on Qwen3-VL-8B-Thinking. It is optimized via LoRA-based RLVR (Reinforced Long-horizon Vision-Language Reasoning) for reliable multi-step execution in agentic automation tasks such as browser and desktop UI control, as well as screenshot-grounded tool calling without error accumulation or drift. This FP8 edition maintains strong multimodal reasoning, structured action generation, and long-horizon stability while enabling more efficient deployment on compatible GPU hardware.

[!important] FP8 8-bit floating point weight and activation quantization using hardware acceleration on GPUs – FP8 W8A8. Quantization W8A8 FP8-dynamic recipe – examples.

Key Highlights

• BF16 · FP8 (F8_E4M3) Compression: Transformer Engine based FP8 quantization reduces VRAM usage and improves inference efficiency while maintaining structured action accuracy.
• Balanced 8B Architecture: Built on Qwen3-VL-8B-Thinking and optimized for stable multimodal reasoning.
• LoRA + RLVR Optimization: Reinforced long-horizon reasoning minimizes drift in multi-step automation tasks.
• Agentic Tool Calling: Generates structured, machine-executable tool calls suitable for automation pipelines.
• Visual Grounding: Accurately interprets UI elements, layout hierarchies, and on-screen context.
• Intent-Driven Actioning: Converts natural language instructions into executable UI actions.
• Cross-Platform Support: Applicable across web apps, desktop software, and mobile interfaces.
• Optimized Deployment: FP8 compression enables efficient inference on supported GPU architectures with reduced memory overhead.

Quick Start with Transformers

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch

# Load the FP8 Jan-v2-VL-med model
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "janhq/Jan-v2-VL-med-FP8",
    torch_dtype="auto",
    device_map="auto"
)

processor = AutoProcessor.from_pretrained(
    "janhq/Jan-v2-VL-med-FP8"
)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "screenshot.png",
            },
            {"type": "text", "text": "Click the settings icon and open the preferences menu."},
        ],
    }
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)

image_inputs, video_inputs = process_vision_info(messages)

inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
).to("cuda")

generated_ids = model.generate(**inputs, max_new_tokens=256)

generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]

output_text = processor.batch_decode(
    generated_ids_trimmed,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False
)

print(output_text)

Intended Use

• GUI Automation Research: Studying multimodal grounding in software environments.
• Agentic Systems Development: Building structured tool-calling pipelines.
• UI-Based Question Answering: Extracting information from complex interface layouts.
• Web and Desktop Automation: Executing structured action sequences in live applications.
• Human-Computer Interaction Research: Evaluating intent-to-action alignment in multimodal agents.

Limitations & Risks

Important: This model generates executable structured actions.

• Execution Risk: Generated tool calls may trigger real actions if connected to live systems. Proper validation layers are strongly recommended.
• Context Sensitivity: Performance depends heavily on screenshot clarity and UI complexity.
• Hardware Requirements: FP8 requires compatible GPU hardware support for optimal performance.