Qwen 3.8 Minecraft benchmark: what 120 FPS means

In an account published October 25, 2026, Adam Conway reported that Minecraft ran at around 120 frames per second for much of a session while Qwen 3.8 Flash Next ran locally on an HP OmniBook Ultra 16 with RTX Spark. Qwen chose actions for a bot in a live Minecraft world; software called Mindcraft and Mineflayer carried them out.

Qwen chose actions, while Mindcraft handled the game

The test connected Qwen to a Minecraft world over a local network. The model selected and narrated what the bot should do. Mindcraft, using its Mineflayer dependency, handled movement, mining and inventory management. The result was an AI-directed agent, not a model creating a Minecraft game.

The local setup served Qwen through llama.cpp, which exposed an OpenAI-compatible API on port 9931. Mindcraft’s LM Studio integration connected to that server. Minecraft ran on Windows on Arm with Arm64 Java and graphics libraries, without x86 emulation through Microsoft PRISM.

The system was described as having a 20-core Arm CPU, a 6,144-core Blackwell GPU and 128GB of unified memory. Qwen 3.8 Flash Next was described as a 125-billion-parameter mixture-of-experts model, a design that activates selected parts of a model for each token. About 6 billion parameters were active per token, while a 51-billion-parameter n-gram embedding was streamed from storage.

Gathering worked; navigation was less reliable

The bot gathered birch logs and dropped eight saplings, but it also made inventory mistakes, got trapped and disconnected. Conway attributed the navigation problems to Mineflayer’s pathfinding and Mindcraft’s recovery behavior.

Mindcraft’s recovery logic detects 20 seconds without movement, tries to move the bot five blocks, and terminates and retries the agent if it remains stuck for another 10 seconds. Those failures concerned the bot’s navigation and recovery, not the reported game frame rate.

What the reported 120 FPS means

The around-120-FPS figure applies to much of Conway’s one Minecraft session with local inference running at the same time. Conway cautioned that a more demanding game could fare worse. He also reported that RTX Spark did not support CUDA in Graphics, so the game and model shared GPU resources without that feature.

Before launching Minecraft, Conway measured Qwen’s inference at roughly 46 tokens per second with a short context and just under 42 tokens per second with a 32,000-token context. Those figures measure generated tokens, not game frames, and were recorded before Minecraft started.

Controlling a game and generating one are different tasks

A separate coding challenge asked models to create a functional Minecraft-style demo in one self-contained HTML5, CSS and JavaScript file. That is a code-generation task, unlike choosing actions for a bot inside a live Minecraft world.