Laguna S 21 Goes Greek

Laguna S 2.1 goes Greek: a repository-scale game transformation Try the new DeepSeek V4 Flash today. Frontier intelligence at a fraction of the cost. Here

AI models

Laguna S 2.1 goes Greek: a repository-scale game transformation

We put Poolside’s new Laguna S 2.1 model to the test, tasking it with a repository-scale transformation of the open-source game Hypersomnia.

Authors

Albert Lee

Last updated August 3, 2026

Share

TL;DR We tested a repository-scale transformation to see whether an agent could understand an unfamiliar codebase, coordinate changes across files and tools, and continuously verify its work. Poolside Laguna S 2.1 , now available on Baseten Dedicated Inference, took on the challenge with a mythological twist.

We gave Poolside Laguna S 2.1 a simple task: transform Hypersomnia , a free and open-source multiplayer top-down shooter game, from neon cyberpunk style into an Ancient Greek odyssey. To do this, Laguna had to understand a complex codebase, coordinate work across multiple AI models, and continuously verify its changes to ensure the gameplay was intact. This post follows Laguna from its first pass through the repository to the final playable build. Along the way, we examine how it orchestrated specialized models, scaled generated assets across hundreds of sprites, preserved the game’s technical constraints, and decided when not to use an available tool. Meet Laguna S 2.1 Laguna S 2.1 is Poolside’s latest model for agentic coding and long-horizon work: 118B-parameter Mixture-of-Experts architecture, with 8B parameters activated per token

Up to a 1 million token context window

Thinking and no-thinking modes

In Poolside’s published evaluations , Laguna S 2.1 performed competitively with substantially larger models on agentic coding benchmarks.

Putting benchmarks to the test: transforming an entire repository Benchmarks can measure coding and tool use, but they cannot fully show whether a model can navigate an unfamiliar codebase, make creative decisions, and recover when an approach fails. This repository-scale transformation put those capabilities to a more practical test. Before making any changes, Laguna had to understand the codebase’s architecture, dependencies, asset pipeline, and undocumented assumptions. It then had to: Coordinate changes across interdependent files.

Call the right tools and external models at the right time.

Build and test continuously, diagnosing failures rather than merely detecting them.

Revise the plan without drifting from the original task.

The Hypersomnia codebase is built from scratch in modern C++ without a commercial engine; the open-source game spans 715 compiled source files that power everything from gameplay and networking to physics, visual assets, and an in-game map editor. That scale turned a playful, creative brief into a tightly constrained engineering task. Laguna had to transform the game’s neon cyberpunk identity into an Ancient Greek odyssey by modifying code, in-game text, configuration, and hundreds of visual assets. And it managed to do it without changing its simulation, geometry, networking, or core gameplay. Unlike a bounded bug fix, this task combined creative ambiguity with technical interdependence across an entire repository. Testing Laguna as a model orchestrator To make the test even more compelling, we added another wrinkle. Rather than ask Laguna to do everything itself, we deployed three supporting models on Baseten Dedicated Inference : Step 3.7 Flash , a fast vision-language model built for high-volume agentic work

Krea 2 Turbo , a text-to-image model built for fast, high-quality image generation

NVIDIA Cosmos 3 Nano , an 8B world foundation model built for text-to-video and image-to-video generation

Running them on dedicated deployments meant Laguna could call them as often as the task required without budget constraints, but Laguna also had to choose the right model for each job, turn its output into something the repository could use, and keep every contribution from breaking the game. ✕ Laguna acts as a model orchestrator The results After 98 minutes and 58 seconds of autonomous work in pool , Poolside’s coding agent, Laguna S 2.1 produced “Hyperion,” an Ancient Greek-themed proof of concept that successfully launched Hypersomnia’s single-player tutorial. ✕ Hypersomnia became Hyperion: Laguna replaced the game’s cyberpunk aesthetic with an Ancient Greek visual identity while preserving the gameplay underneath. The run at a glance: Laguna token usage: 76.33M input tokens and 449.5K output tokens

Context used: 590.5K of 1.05M tokens (56%)

Supporting model usage: 15 vision-language model calls (Step 3.7 Flash) and 100 image-generation requests (Krea 2 Turbo), with 53 successful image generations

Assets modified: 954 PNG sprites, composited from 53 generated images

Code and config files modified: 34 .cpp, 15 .h, 7 .hpp, and 8 JSON files

The result was playable, but not production-ready. Hyperion could not connect to existing Hypersomnia multiplayer servers. It did launch the single-player tutorial successfully, confirming that the new assets worked with the core gameplay. ✕ Laguna S 2.1 produced a playable Greek-themed single-player tutorial. Laguna’s creative judgment in practice The playable tutorial was the final frame of a more layered story where delegation, revision, and a careful balance between creative ambition and technical restraint helped shape the path to get there. That judgment surfaced in three ways: Laguna directed Step’s language work, scaled Krea’s visual outputs, and recognized that Cosmos had no role in the game itself. Step 3.7 Flash: mythic language, familiar meaning Laguna tasked Step 3.7 Flash with adapting the game’s language to the Ancient Greek theme. It specifically prompted Step to “produce a Greek-themed replacement with playful absurdity, NO profanity, kept SHORT and understandable.” Rather than applying Greek names at random, Step translated each term according to its role in the game. Factions kept their power relationships, character classes remained recognizable, and objectives and game modes preserved their original meaning even as the language became more mythic.

Laguna remained the final editor. When Step returned Greek-language copy that would be inaccessible to most players, Laguna rewrote it in clear English while preserving the mythic tone. Its role went beyond delegation, and it judged whether each output served the larger...