How to use from the
Use from the
Transformers library
# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="prithivMLmods/KORMo-Web-4B-MTP-Repacked")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)
# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("prithivMLmods/KORMo-Web-4B-MTP-Repacked")
model = AutoModelForMultimodalLM.from_pretrained("prithivMLmods/KORMo-Web-4B-MTP-Repacked", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))
Quick Links

KORMo-Web-4B-MTP-Repacked

KORMo-Web-4B is an open-weight 4B bilingual (Korean and English) vision-language web agent developed by KAIST researchers on top of Qwen3.5-4B that executes complex, multi-site, and multi-tab browser tasks purely from screenshotsβ€”requiring no DOM or accessibility tree. Pretrained on over 10 million atomic-level web element understanding samples and fine-tuned on more than 100,000 real browser interaction trajectories, it uses a step-by-step <think> reasoning loop across 22 normalized-coordinate browser actions (click, type, scroll, tab control, etc.) to achieve a 42.3% success rate on both English and Korean Online-Mind2Web benchmarks (+22.3 and +28.8 points over its base model, outperforming comparable open-weight agents). Released under an Apache 2.0 license alongside a Playwright reference harness and a Chrome extension, it runs on a single NVIDIA GPU via vLLM or locally on Apple Silicon Macs via vllm-mlx (with a 4-bit variant fitting in 16 GB RAM).

MTP β€” Speculative Decoding: Yes | GGUF β€” KORMo-Web-4B-MTP-Repacked(GGUF)

Multi-Token Prediction (MTP) GGUF is a specialized GGUF model file format extension that integrates speculative decoding directly into the model weights to significantly accelerate local inference. Unlike traditional speculative decoding which requires a separate, smaller "draft" model, MTP GGUF files include additional output heads within the main model architecture that predict multiple future tokens in a single forward pass.

Model Files [GGUF]

File Name Quant Type File Size File Link Description
KORMo-Web-4B-MTP.BF16.gguf BF16 9.94 GB Link Full BF16 weights. Highest quality, largest file size.
KORMo-Web-4B-MTP.Q3_K_L.gguf Q3_K_L 2.82 GB Link Lower quality but usable, good for low RAM availability.
KORMo-Web-4B-MTP.Q3_K_M.gguf Q3_K_M 2.63 GB Link Low quality, smallest file size.
KORMo-Web-4B-MTP.Q4_K_M.gguf Q4_K_M 3.17 GB Link Good quality, default size for most use cases, recommended.
KORMo-Web-4B-MTP.Q4_K_S.gguf Q4_K_S 2.99 GB Link Slightly lower quality with more space savings, recommended.
KORMo-Web-4B-MTP.Q5_K_M.gguf Q5_K_M 3.62 GB Link High quality, recommended.
KORMo-Web-4B-MTP.Q5_K_S.gguf Q5_K_S 3.51 GB Link High quality with slight space savings, recommended.
KORMo-Web-4B-MTP.Q6_K.gguf Q6_K 4.08 GB Link Very high quality, near perfect, recommended.
KORMo-Web-4B-MTP.Q8_0.gguf Q8_0 5.29 GB Link Extremely high quality, generally unneeded but max available standard quant.
KORMo-Web-4B-MTP.mmproj-bf16.gguf mmproj-bf16 676 MB Link Multimodal projection file in BF16 format. Used for vision/image capabilities.

MTP-Repacked [Transformers, BF16]

File Name File Type File Size File Link Description
chat_template.jinja Jinja Template 7.76 kB Link Chat template for formatting conversations and prompts.
config.json JSON 2.85 kB Link Core model architecture and configuration parameters.
generation_config.json JSON 219 Bytes Link Default text generation parameters (e.g., temperature, top_p).
model-00001-of-00002.safetensors Safetensors 5.49 GB Link Model weight shard 1/2. Required to load the model.
model-00002-of-00002.safetensors Safetensors 5.1 GB Link Model weight shard 2/2. Required to load the model.
model.safetensors.index.json JSON 67.4 kB Link Index file mapping model weights to their respective shards.
processor_config.json JSON 1.19 kB Link Configuration for the multimodal processor.
system_prompt.md Markdown 3.28 kB Link Default system prompt or behavioral guidelines for the model.
tokenizer.json JSON 20 MB Link Complete tokenizer vocabulary and encoding rules.
tokenizer_config.json JSON 1.16 kB Link Tokenizer-specific configuration parameters.
tools.json JSON 12.4 kB Link Configuration or definitions for available tools/functions the model can use.

llama.cpp

LLM inference in C/C++ β€” https://github.com/ggml-org/llama.cpp

MTP-repackaged and quantized using llama.cpp by β€” prithivMLmods.

Downloads last month
434
Safetensors
Model size
5B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for prithivMLmods/KORMo-Web-4B-MTP-Repacked

Finetuned
Qwen/Qwen3.5-4B
Quantized
(3)
this model

Collections including prithivMLmods/KORMo-Web-4B-MTP-Repacked