Instructions to use itapitarules/OpenCore-1M-MTP-4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use itapitarules/OpenCore-1M-MTP-4B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="itapitarules/OpenCore-1M-MTP-4B")
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)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("itapitarules/OpenCore-1M-MTP-4B")
model = AutoModelForMultimodalLM.from_pretrained("itapitarules/OpenCore-1M-MTP-4B", 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=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use itapitarules/OpenCore-1M-MTP-4B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "itapitarules/OpenCore-1M-MTP-4B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "itapitarules/OpenCore-1M-MTP-4B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/itapitarules/OpenCore-1M-MTP-4B

SGLang

How to use itapitarules/OpenCore-1M-MTP-4B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "itapitarules/OpenCore-1M-MTP-4B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "itapitarules/OpenCore-1M-MTP-4B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "itapitarules/OpenCore-1M-MTP-4B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "itapitarules/OpenCore-1M-MTP-4B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use itapitarules/OpenCore-1M-MTP-4B with Docker Model Runner:
```
docker model run hf.co/itapitarules/OpenCore-1M-MTP-4B
```

OpenCore 1M MTP 4B

This is an experimental candidate unquantized BF16 OpenCore 2BPlus2B edition. It is one standard Qwen3.5 multimodal model, not two models glued together. Every released parameter is on a normal executable language, vision, or MTP path. The coding specialization is merged into the language matrices. No workflow system prompt is injected at runtime.

Exact contents

Total serialized parameters: 4,333,965,344
Language parameters: 3,941,720,608
Vision parameters: 331,416,576
Native one-block MTP parameters: 60,828,160
Main plus vision-projector GGUF size: 8.690 GB
Tensor types: BF16 plus only architecture-required F32 auxiliaries; no quantization
Selectable context metadata: 1 to 1,010,000 tokens
Original trained context: 262,144 tokens
Measured selected MTP draft maximum: 4
Measured local median generation speed: 53.8464 tokens/second
Coding/tool evaluation: exact executable task success 50%, tool validity 100%; candidate rejected

The million-token value is the loader-selectable ceiling. Memory use grows with the context selected, so start at 32K on a 12 GB GPU and increase only when the task needs it. The original trained window is 262,144 tokens; longer selections are extrapolation and are not a claim of equal quality across the full ceiling.

Files

OpenCore-1M-MTP-4B-BF16.gguf: llama.cpp / LM Studio main BF16 GGUF, including native MTP tensors
OpenCore-1M-MTP-4B-BF16-mmproj.gguf: BF16 multimodal vision projector
OpenCore-1M-MTP-4B-BF16-HF/: exact merged Hugging Face BF16 model
Modelfile: Ollama text-mode import, with no hidden SYSTEM prompt
run-opencore.ps1: llama.cpp launcher using GPU offload, Flash Attention and MTP
RELEASE_INDEX.json: checksums, sizes, runtime status and architecture facts

Run

LM Studio: import the main GGUF, select its matching projector, and choose context at load time. llama.cpp from this folder:

.\run-opencore.ps1 -Context 32768 -DraftMax 4

Ollama text-mode import:

ollama create opencore-1m-mtp-2bplus2b -f Modelfile

Ollama backend support for external multimodal projectors and native Qwen3.5 MTP may differ from llama.cpp. The full-feature validation path is recorded in the release index.

Real cross-harness learning

A GGUF file is immutable and cannot safely rewrite itself during generation. OpenCore's companion stores only objectively successful interactions in the shared SQLite store, trains a candidate, runs protected quality gates, rebuilds a fresh BF16/GGUF generation, and atomically promotes it only when it does not regress. A rejected or failed build leaves the current model untouched and supports rollback. The edition config is configs\continual-2bplus2b.json.

python scripts\background_learn.py --gguf-config configs\continual-2bplus2b.json

The detailed measured reports are artifacts\2bplus2b-speed-mtp4-recheck.json and artifacts\2bplus2b-interleaved-cp120-fresh-v11-exact.json.

Downloads last month: 47

Safetensors

Model size

4B params

Tensor type

F32

BF16