MiniMax H3 Can Now Be Steered Like a Game, Using Barely Any New Training

A new research project called H3-World turns MiniMax H3, the 33-billion-parameter video and audio model BRC has covered extensively, into a controllable, game-like simulator. Press W, A, S, or D and a character in the generated scene moves accordingly. Press I, J, K, or L and the camera responds. The genuinely striking part is how little new training this required: just 0.199% of the base model's parameters.

How H3-World Actually Works

Researchers Danze Chen, Zeqing Wang, Ziyue Lin, Xingyi Yang, and Yeying Jin built H3-World as a LoRA adaptation sitting on top of MiniMax H3's existing weights, rather than a new model trained from scratch. The system converts each keyboard input into a language instruction, one per future video segment, and uses what the paper calls "directed attention routing" to bind that instruction to the correct interval of generated video. In effect, it's teaching the model to translate discrete key presses into the same kind of natural-language scene direction it already understands from text prompts, rather than building an entirely separate control system.

The training itself was small by modern AI standards: 65.6 million LoRA parameters, trained on just 8,000 gameplay clips from a dataset called ABot-World-Explorer-500h. That's a tiny fraction of the data and parameter count typically associated with building a capable interactive model, evidence that H3's underlying multimodal understanding of language and video was already rich enough to repurpose rather than retrain.

What You Actually Get

Given a starting frame and text description, H3-World generates action-controlled video where character movement (WASD) and camera motion (IJKL, with F for fast movement) both respond to live input, producing coordinated motion and camera work rather than the character and camera behaving independently.

The released checkpoint requires the base MiniMax-H3 weights plus H3-World's specific attention-routing patch; it won't reproduce the reported behavior through an unmodified H3 pipeline. The LoRA checkpoint itself is released under Apache 2.0, though the underlying MiniMax-H3 weights remain governed by their own separate license.

Competitive Context

Interactive, keyboard-controllable world models are a genuinely active research category right now, Nvidia's Cosmos, Wayve's GAIA-2, and projects like Matrix-Game and ActWorld are all working similar territory, mostly aimed at gaming, robotics, and autonomous driving simulation.

What distinguishes H3-World within that field isn't the concept, it's the method: rather than building and training a dedicated world model architecture from the ground up, this project demonstrates that an existing large multimodal video model can be steered into that behavior with a comparatively tiny, efficient adaptation layer.

The Signal in the Noise

The efficiency story here matters more than the demo itself. If a general-purpose video model's existing language understanding can be redirected into real-time interactive control using under 0.2% additional trained parameters and a relatively small gameplay dataset, that's a meaningfully lower barrier to building interactive or game-like AI tools than training a dedicated world model from zero.

It's a strong signal that a lot of the "hard part" of interactive world modeling may already live inside general-purpose multimodal models, waiting to be unlocked rather than built anew.

Does knowing how little retraining this required change how you think about what's actually possible to build on top of models like MiniMax H3, beyond straightforward video generation?

The Details

• Project: H3-World, built on MiniMax H3 (33B parameters)
• Method: LoRA adaptation with directed attention routing, not a full retrain
• Trained parameters: 65.6M, 0.199% of the base model
• Training data: 8,000 gameplay clips from ABot-World-Explorer-500h
• Controls: WASD (character movement), IJKL (camera), F (fast camera motion)
• Requirements: base MiniMax-H3 weights plus H3-World's attention-routing patch
• License: H3-World LoRA checkpoint under Apache 2.0; MiniMax-H3 base weights under separate license
• Paper: "H3-World: Turning Language Understanding into World Control," arXiv 2609.01560

Resources & Reads

• H3-World on Hugging Face
• H3-World research paper on arXiv
• BRC: How fal Turned an Open-Weight Model Into the Fastest AI Video Tool Available