Ever wondered what would happen if the world's leading Large Language Models settled their benchmark disputes in a 2D cybercity arena?
It's easy to look at model performance on standardized benchmarks (like MMLU, MATH, or HumanEval). It is much more fun to visualize their underlying architectures, parameter scales, and hardware constraints as a retro-cyber fighting game.
So, we built Model Kombat (Mixture of Experts Edition)!
๐น๏ธ Play Directly Here
๐ฎ Launch Game in Full Screen
๐งฌ Playable ML Concepts Explained
This isn't just a basic stick-figure fighting game. Every mechanicโfrom rendering complexity to the speed at which characters recoverโis a direct, playable representation of real-world Large Language Model engineering.
1. ๐ Parameter Scaling vs. Render Tiers
A model's representation capacity (intelligence) scales with its parameter count. In Model Kombat, a fighter's visual complexity, joint detail, and rendering fidelity directly reflect its real-world parameter size:
- Tier 1 (< 5B Parameters - Gemma 2B, Llama 3.2 3B) - Primitive Capsules: Drawn as simple, single-color flat limbs with low joint segmentation. This visualizes the limited representation capacity and coarse output resolution of small edge models.
- Tier 2 (7B - 14B Parameters - Mistral 7B, Claude Haiku) - Simple Vectors: Structured as thin skeletal wireframe vectors.
- Tier 3 (14B - 35B Parameters - Gemini Flash, Mixtral) - Two-Tone Vectors: Rendered as dual-color, layered vector limbs.
- Tier 4 (35B - 100B Parameters - Llama 8B, Claude Sonnet) - Cyborg Shading: Rendered as detailed vector cylinders with dynamic code particle streams flowing along their limbs.
- Tier 5 (> 100B Parameters - o3, GPT-4o, Claude Opus) - Quantum Vectors: Rendered as glowing vector limbs with digital matrix code particles, soft drop-shadow depth buffers, and real-time afterimage motion trails.
2. โก Reasoning Tokens & KV-Cache Overcharging
Instead of arbitrary "mana" or "stamina," fighters charge a Ki bar representing internal processing cycles and Reasoning Tokens (inspired by reasoning chains like OpenAI's o-series):
- Charging Ki: Simulates the time-to-first-token (TTFT) phase, generating reasoning tokens.
- Limit Break: Overcharging past 100% enters a golden-outlined Limit Break state, granting high-speed afterimages and super-armor.
- Context Eviction & Dizzy: If a model holds its overcharged state too long, its context window overflows. This triggers Context Evictionโdraining the model's HP and placing it in a Dizzy state. This represents how context window saturation degrades model coherence and leaves it vulnerable to failure.
3. ๐ Mixture of Experts (MoE) Routing
Sparse Mixture of Experts (MoE) models do not activate all parameters on every token; instead, a gating network routes tokens to specialized experts.
- Active Experts: Native MoE models in the game (like Mixtral and DeepSeek) dynamically spawn floating indicator nodes representing active Text, Math, or Vision experts.
- Routing Buffs: Landing hits routes computation to these experts, granting temporary combat buffs:
-
TEXTFFN: Increases walk speed (low-latency generation). -
MATHFFN: Increases damage output (logical compute). -
VISIONFFN: Expands attack hitboxes (spatial awareness).
-
๐ Spotlight: The AI Fighter Registry
Here is a breakdown of 6 key models featured on the ladder, explaining where they excel in real life and how their unique passives translate into combat advantages:
1. ๐ง o3 (OpenAI)
- In Real Life: OpenAI's state-of-the-art reasoning model. Rather than generating text instananeously, o3 employs a reinforcement learning-driven "thinking chain" to plan, verify, and correct its logic before outputting a response. This makes it a titan in mathematics, competitive programming, and complex coding.
- In-Game (Tier 5): Deep Thinking Chain passive. o3 charges its Ki at double speed (representing the model's heavy pre-response thinking cycles). This lets you quickly max out your meter, activate MoE, or release powerful special attacks.
2. ๐ Claude Opus (Anthropic)
- In Real Life: Anthropicโs flagship heavy model. Opus is celebrated for its high-nuance reading comprehension, literary synthesis, and strict compliance with ethical and safety guidelines (governed by Anthropic's "Constitutional AI" framework).
- In-Game (Tier 5): Constitutional Blade passive. Opus has an extended melee strike range on all punches and kicks, allowing you to control the neutral game and keep opponents at a distanceโvisualizing the model's massive context processing reach.
3. ๐ค Gemini Ultra (Google)
- In Real Life: Google's largest multimodal model. Unlike models that stitch together separate speech-to-text and vision encoders, Gemini is built natively multi-modal from day one. It processes video, audio, and text simultaneously inside a single model architecture.
- In-Game (Tier 5): Ultra Stance Shift passive. Gemini can switch between
TEXT,VISION(expanded hitboxes), andAUDIO(increased speed) modality stances instantly without the standard stance-transition delay, adapting to any opponent's position on the fly.
4. โก DeepSeek V3 (DeepSeek)
- In Real Life: The groundbreaking open-weights model from DeepSeek. V3 utilizes Multi-Head Latent Attention (MLA) to compress Key-Value caches, dramatically reducing VRAM footprint, alongside a massive Multi-head Latent Attention routing gating network.
- In-Game (Tier 4): MLA Attention passive. Compress caching translates to high combat evasion: DeepSeek V3 has a 15% chance to phase-dodge incoming projectiles completely.
5. ๐ผ Mistral 7B (Mistral)
- In Real Life: The legendary French open-weights model that punches far above its weight class. Mistral 7B introduced Sliding Window Attention (SWA) to the open-source community, allowing the model to handle longer context streams with minimal performance decay.
- In-Game (Tier 2): Sliding Window passive. The sliding window translates to faster execution: Mistral has 10% less startup frame lag on punches and kicks, letting you land strikes before your opponent's animations can finish.
6. ๐ฑ Llama 3.2 3B (Meta)
- In Real Life: Metaโs mobile-first, edge-optimized model. Llama 3.2 3B is trained specifically for local deployment on smartphones and tablets, focusing on high efficiency, quick response times, and localized fine-tuning.
- In-Game (Tier 1): Fine-Tuning passive. The model is incredibly resilient to pressure: Llama 3.2 gains +5% defense every time it successfully blocks consecutive hits, adapting to the opponent's combo string.
๐ Beat the Machine & Share Your Score
Once you defeat all 19 opponents on the ladder, you will face the reigning champion: o3.
Can you reach the top of the ladder and claim victory? Click COPY SCORE at the end of your run and paste your stats in the comments below!
๐ฌ Let's Discuss:
- Which model is your favorite main?
- What was your longest combo chain on the ladder?
UnitBuilds-CC
/
Model-Kombat
Arcade fighting game, which LLM will come out on top?
๐ Model Kombat (SYS_08) ๐ง ๐ฅ
An Interactive, Playable Visualization of Large Language Model Constraints & Architectures
Welcome, AI engineer. Model Kombat is a retro-cyber fighting game designed to teach the mechanical constraints, architectural paradigms, and hardware limits of Large Language Models (LLMs).
In this game, 20 real-world AI models face off in a 2D arena. Every visual element, movement mechanic, and combat stance directly translates core machine learning engineering concepts (such as parameter scaling, KV-caching, Mixture of Experts, multi-modality, and alignment safety) into interactive gameplay loops.
๐น๏ธ Play Directly Here
๐ฎ Model Kombat Live Web App
- Git Repository: https://github.com/UnitBuilds-CC/Model-Kombat
๐ง Educational Core Concepts & Game Translations
1. ๐ Parameter Scaling vs. Render Tiers
In deep learning, a model's representation capacity (intelligence) scales with its parameter count. In Model Kombat, a fighter's visual complexity, joint detail, and rendering fidelity directly reflect its real-world parameter size:
- Tier 1 (< 5B Parameters -โฆ
Disclaimer: AI was used throughout this project, it is just fitting that it would co-author with me, so special thanks to the Foundry for its tireless hours toiling away and Gemini for producing the cover image.
Top comments (19)
Awesome, keep going!
So how do I win? The enemy has no HP left but just won't go down.๐คฃ
And I have to say โ your design skills are an absolute disaster. ๐
I'm probably gunna end up redo'ing all the games this weekend, I decided to build a proper game engine, etc. So I dont have to write 3000+ lines per game, with it, things are alot more robust, so I'll quickly hot-patch, then redo it properly via the game engine, so it actually plays better. Cuz imo, Model Kombat is pretty meh, it needs ALOT of work, which the game engine will hopefully help with
Go get 'em! May the Colonel be with you. ๐
Give it a try now, patch deployed
Still the same๐คฃ โ is it a cache thing๏ผ
Probably. Cuz when I test it my side, it works. It's supposed to throw a 'FINISH HIM' and executes the model's finisher move to do a Fatality
Damn... First production bug. I'll look at what's up and get it fixed asap. Thanks for letting me know!
Broo! This is a good one! But you can probably upgrade it too...! ๐ Anyways, Nice work! ๐๐
It's gunna get a full overhaul. I spent most of yesterday building out a full game-engine, workflows, etc. cuz I got tired of having to rebuild everything from scratch daily.
No worries! Take your own time! Wishing you all the best ๐โค๏ธ!
Updated pushed and tested. Though not perfect, I'll need to first build out the more complex systems of the engine before I can further improve it. Hope it atleast plays a bit better now
Game-Gen incase anyone is interested in the progress of the engine
Question. At the moment, making these games are time consuming, I build each from scratch, reusing almost nothing... Do I just create a lightweight game-engine for it? Would you all be interested in an open-source royalty free game-engine?
Hi. How about we discuss the path to becoming wealthy through reasonable cooperation?
This game is a lot of fun, and the animation is outstanding. How long did it take you to make it?
Thank you! In total, probably around 8h, wherever I had time to work on it. Though now that i have the game engine up, should be able to cut that in half.
@xulingfeng Satisfied? ๐
@xulingfeng never Satisfied!