tinywmcraft

tinywmcraft is an autoresearch benchmark and 14M-parameter world model of Craftax trained from pixels and actions.

Tinywmcraft is a 14M-parameter world model of Craftax-Classic, a 2D Minecraft game, learned from scratch using pixels and actions alone. Given a history of frames and an action, the system generates the next frame, the reward, and whether the episode ends. Users can feed the generated frames back in to play inside the browser, where every world starts from an initial frame imagined from noise. It is designed for researchers and coding agents to train, benchmark, and push scores up autoresearch style. The model consists of 13.9 million parameters split between a tokenizer and a world model, and trains cold in under 10 minutes on a single A100 GPU.

tinywmcraft at a glance

Best for: Software engineer
Runs on: Web
Pricing: Colab Pro $9.99 for 100 compute units; training a full submission costs under $0.5.
Released: November 2026
Category: AI research tool
Website: maharishiva.github.io

Key features

Train a 14M-parameter world model
Play inside the world model in the browser
Autoresearch benchmark for coding agents
Score models based on image, logic, and reward prediction
Generate next frames, rewards, and episode ends
Initialize worlds from frames imagined from noise

Use it for

Train a world model of Craftax in under 10 minutes
Play inside an AI-generated 2D Minecraft world
Compete in a diamond speedrun challenge
Direct coding agents to train and score world models

Alternatives to tinywmcraft

AntV Chart MCP — Provides direct integration of diverse chart generation tools into AI applications.
Simulusion — Compare and train multiple world models using one shared trainer and config file.
sizeless — Generates centimeter-accurate 3D point clouds and CAD plans directly from iPhone videos without GPS or…
MiroFlow — Achieves top rankings on multiple agentic benchmarks with modular architecture.
Flint — Compiles compact specifications into native output for multiple chart libraries.