Instructions to use EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3 with Transformers:

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

pipe = pipeline("image-text-to-text", model="EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3")
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("EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3")
model = AutoModelForMultimodalLM.from_pretrained("EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3", 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 EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3",
		"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/EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3

SGLang

How to use EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3 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 "EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3" \
    --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": "EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3",
		"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 "EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3" \
        --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": "EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3",
		"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 EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3 with Docker Model Runner:
```
docker model run hf.co/EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3
```

Minecraft Agent — Qwen3.5-9B Stage3 (Text-Action)

A vision-language agent for Minecraft that plays from raw first-person pixels and outputs structured text actions. Fine-tuned from Qwen3.5-9B (stage2 → stage3) on 200k+ human gameplay trajectories (6M frames, 20 FPS, 640×360) converted to a text-action format.

Recipe (Stage 3)

Component	Value
Base	Qwen3.5-9B (stage2)
Data	`minecraft-text-action-dataset-noop-filtered` — 92.85% of the raw dataset after removing 7.15% pure no-op collapse trajectories
`KEEP_NO_OP_P` (no-op frame drop)	0.2 (drop 80% of no-op frames at data level)
Focal loss	decay = 0.75 (down-weights repeated consecutive actions)
Steps	3000 (this checkpoint)

Evaluation

Benchmark: easy-ng — 202 non-GUI tasks (153 Embodied mine_block + 49 Combat kill_entity) × 3 rollouts, difficulty = easy, 200-step episodes, seed=42.

Inference history window	Success rate	Embodied	Combat
3 frames (default)	15.0%	17.9%	6.2%
29 frames (matches training context)	24.9%	24.8%	25.2%

GUI tasks (crafting/smelting/interaction) are out of scope: all text-action models score 0% on them (no interface-coordinate grounding in this action space).

Action format

The model consumes a sliding window of frames (recommend 29 history + 1 current) and emits:

Action: move(dx, dy) and press(keys...) [and click(button)]

e.g. Action: move(0.5, -0.3) and press(w), Action: move(0, 0) and press() and click(left).

The first user turn carries the system prompt + task instruction; subsequent turns are image-only user messages alternating with assistant actions.

Usage

# vLLM (tested with vllm 0.8.5, transformers >= 5.15 required for Qwen3.5)
vllm serve EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3 \
  --limit-mm-per-prompt '{"image": 30}' \
  --enable-prefix-caching \
  --max-model-len 32768

from transformers import AutoModelForImageTextToText, AutoProcessor
model = AutoModelForImageTextToText.from_pretrained(
    "EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3", torch_dtype="bfloat16"
)
processor = AutoProcessor.from_pretrained("EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3")

Notes

Checkpoint: checkpoint-3000 of run minecraft-sft-stage3-qwen35-9b-nf-noop2 (single seed)
Optimizer states removed (weights only, single model.safetensors)
The 29-frame window result exploits the 30-step trajectory segments used in training; enable prefix caching to keep rollout speed comparable to short windows

Downloads last month: -

Safetensors

Model size

9B params

Tensor type

BF16

Model tree for EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Qwen/Qwen3.5-9B

Finetuned

(741)

this model

EmberJin
/

Minecraft-Agent-Qwen3.5-9B-Stage3