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

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 7, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3 is a 9-billion parameter vision-language agent fine-tuned from Qwen3.5-9B for Minecraft gameplay. It processes raw first-person pixels and outputs structured text actions, specifically optimized for embodied and combat tasks within the game. Trained on over 200,000 human gameplay trajectories, this model excels at generating precise actions like movement and key presses based on visual input. Its primary application is autonomous agent control in Minecraft environments, demonstrating a 24.9% success rate on non-GUI tasks with a 29-frame inference history window.

Loading preview...

Minecraft Agent: Qwen3.5-9B Stage3

This model, developed by EmberJin, is a 9-billion parameter vision-language agent specifically designed for playing Minecraft. It is fine-tuned from Qwen3.5-9B (Stage 2 to Stage 3) and processes raw first-person pixel data to generate structured text actions.

Key Capabilities & Features

  • Vision-Language Agent: Operates directly from raw first-person pixel input (640x360 resolution).
  • Text-Action Output: Generates structured text commands like move(dx, dy) and press(keys...) [and click(button)].
  • Specialized Training: Fine-tuned on over 200,000 human gameplay trajectories (6 million frames) converted into a text-action format.
  • Optimized for Embodied & Combat Tasks: Achieves a 24.9% success rate on the easy-ng benchmark (202 non-GUI tasks), including mine_block and kill_entity scenarios.
  • Contextual Awareness: Performance significantly improves with a 29-frame inference history window, matching its training context.
  • Focal Loss: Utilizes focal loss with a decay of 0.75 to down-weight repeated consecutive actions during training.

Use Cases & Limitations

  • Ideal for: Autonomous agent control in Minecraft for tasks involving movement, interaction with blocks, and combat.
  • Not suitable for: GUI-based tasks (crafting, smelting, inventory management) as the current action space lacks interface-coordinate grounding, resulting in 0% success on such tasks.

Finetunes
1 models