Minecraft Agent Qwen3.5-9B Stage 3 GGUF

This repository provides GGUF quantizations and multimodal vision projector weights for EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3.

Minecraft Agent is an embodied vision-language agent trained on 200,000+ human gameplay trajectories (6 million frames at 20 FPS, 640x360 resolution) to play Minecraft directly from raw first-person pixels and output structured text actions.

All GGUF files were converted from the original safetensors weights using llama.cpp at native BF16 precision, then quantized into standard k-quant variants. The vision projector is provided as a standalone mmproj file for multimodal inference.

---

Available Files and Quantizations

---

Technical Specifications

• Base Architecture: Qwen3_5ForConditionalGeneration (model_type: qwen3_5)
• Text Backbone: qwen3_5_text (32 hidden layers, DeltaNet linear attention + full attention)
• Vision Tower: qwen3_5_vision (27-layer ViT, patch size 16, spatial merge size 2, hidden size 1152)
• Parameters: ~9B active
• Context Length: Up to 32,768 tokens
• Action Space: Continuous mouse movements (dx, dy), discrete key presses, and mouse clicks
• Native Precision: bfloat16
• License: Apache 2.0

---

Action Format and Prompting

The agent consumes a sliding window of frames (recommended: 29 history frames + 1 current frame) and outputs structured text actions:

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

Examples

• Walking forward and turning right: Action: move(0.5, -0.3) and press(w)
• Mining a block in place: Action: move(0, 0) and press() and click(left)
• Jumping while sprinting forward: Action: move(0, 0) and press(w, space)

Turn Structure

The first user turn supplies the system prompt, task instruction, and initial frame. Subsequent turns provide image-only user messages alternating with assistant action predictions.

---

Benchmark Results (easy-ng)

Evaluated on the easy-ng benchmark across 202 non-GUI tasks (153 Embodied mine_block + 49 Combat kill_entity), 3 rollouts per task, difficulty easy, 200-step episodes:

---

How to Use

1. With llama.cpp

For multimodal generation with raw gameplay screenshots, pass both the quantized model and the mmproj file:

llama-qwen2vl-cli \
    -m ./Minecraft-Agent-Qwen3.5-9B-Stage3-Q4_K_M.gguf \
    --mmproj ./mmproj-Minecraft-Agent-Qwen3.5-9B-Stage3-BF16.gguf \
    --image ./gameplay_frame.png \
    -p "Mine the wood block directly ahead." \
    -n 64

For text-only testing or action syntax verification:

llama-cli -m ./Minecraft-Agent-Qwen3.5-9B-Stage3-Q4_K_M.gguf \
    -p "<|im_start|>system\nYou are an embodied Minecraft agent.<|im_end|>\n<|im_start|>user\nMine a tree block in front of you.<|im_end|>\n<|im_start|>assistant\nAction: " \
    -n 32

Launch an OpenAI-compatible multimodal server:

llama-server \
    -m ./Minecraft-Agent-Qwen3.5-9B-Stage3-Q4_K_M.gguf \
    --mmproj ./mmproj-Minecraft-Agent-Qwen3.5-9B-Stage3-BF16.gguf \
    --port 8080 \
    -c 16384 \
    --host 0.0.0.0

2. With Ollama

Create a file named Modelfile:

FROM ./Minecraft-Agent-Qwen3.5-9B-Stage3-Q4_K_M.gguf
MMPROJ ./mmproj-Minecraft-Agent-Qwen3.5-9B-Stage3-BF16.gguf

TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{- end }}
{{- range .Messages }}
<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{- end }}
<|im_start|>assistant
Action: 
"""

PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER temperature 0.2
PARAMETER top_p 0.9

Build and run:

ollama create minecraft-agent -f Modelfile
ollama run minecraft-agent

---

Limitations

• GUI tasks such as crafting tables, furnace smelting, and chest sorting are out of scope (the model operates strictly on pixel-to-motor control and lacks inventory GUI coordinate grounding).
• Optimal execution requires a sliding history of up to 29 frames to maintain behavioral momentum.

---

License

This model is licensed under Apache 2.0, inheriting upstream licensing from Qwen/Qwen3.5-9B. See the LICENSE file for details.