Poolside’s Laguna S model and enterprise codebase migration strategies

Poolside uses the 118-billion-parameter Laguna S model to automate software development tasks. This model uses a sparse Mixture-of-Experts architecture where only 8 billion parameters remain active during processing. It outperforms models two to three times its size on coding benchmarks because it prioritizes persistence, such as backtracking and trying multiple approaches, over raw intelligence. The company built its "Model Factory" from scratch over three years to run 10,000 to 20,000 experiments per month using a team of fewer than 70 researchers and 35 engineers. This infrastructure allows the team to stream data into training via an internal immutable data layer called the "blender." Instead of forking existing open-source repos, the team wrote their own training codebase. The company trains its models on over 500,000 open-source codebases to reach human-level or superhuman intelligence in software engineering. This effort involves building custom models from the ground up. The founders, Jason Warner and Eiso Kant, designed this system to move from a researcher’s idea to a trusted experimental result quickly. The company manages its training through a just-in-time approach that allows jobs to begin while data is still materializing.

Enterprise Scale and Strategic Investment

Nvidia led a strategic investment of $1 billion in August 2026, which valued Poolside at $13 billion post-money and included a separate non-exclusive license for $6 billion. This investment follows a $500 million Series B round in October 2024 that valued the company at $3 billion. The Series B round included participants such as Bain Capital Ventures, DST Global, and HSBC Ventures. Poolside targets Global 2000 enterprises and public-sector agencies by deploying models inside customer environments. While Cognition and Lovable show higher end-user adoption, Poolside lacks the visible end-user adoption seen in its competitors. The company focuses on specialized software engineering for high-compliance environments like government and defense. Developers use the model through the terminal, editor extensions, or third-party agents. The model supports specific tools. These include shell, kill, shell wait, write, fetch web, and bash. The company was founded in 2023 by Jason Warner and Eiso Kant. The company has raised $1.6 billion across four funding rounds. This growth allows the company to compete with firms like Cognition, which reaches a $900 million annualized revenue run rate.

Managing Tokens for Codebase Migrations

Analyzing a multi-file codebase for legacy migrations requires a sufficient context window to keep all relevant code in view. The context window acts as the short-term memory of the AI, measured in tokens. One thousand tokens equal approximately 750 words in English, or about 4,000 characters. If you exceed the limit, the model drops the oldest or least prioritized information. You should budget 20 to 40 percent of the window for the model’s own working space and outputs. Large windows are not a magic bullet because performance can degrade with scale unless the model was specifically trained to handle larger inputs. Tokenization involves engines like Byte-Pair Encoding or WordPiece to convert text into sub-word pieces. Different tokenizers make different choices about how to divide text. They often keep common words intact while splitting rare ones. You can also use retrieval-augmented generation to let the model look things up instead of using only the information it was trained on. You can also use summaries to keep prompts short and help the AI adapt to your needs. Developers can use different strategies to manage token usage. You can use token counters during development to avoid hitting limits. One can also use semantic chunking to divide code by functions or chapters. This method ensures the AI connects different sections together. Can the model maintain coherence when the codebase size scales to millions of tokens?