Work through coding tasks that span an entire codebase

Laguna S 2.1 is a coding agent engine. You hand it a goal, such as fixing a failing test or tracing a bug across modules, and it plans the steps, edits files, calls tools, and keeps thinking between tool calls instead of losing the thread on step three. The headline ability is the context window: about a million tokens, enough for a full mid-size codebase plus its docs, logs, and your conversation, all in one session. Questions like "where does this value actually get set?" stop requiring you to guess which file to paste.

The trick that makes it practical is its mixture-of-experts design. The model stores 118 billion parameters of knowledge, but only about 8 billion of them wake up for any given word. In daily use that means it reads, thinks, and bills like a small model while drawing on a much larger one. On coding tests the tradeoff holds up: 70.2 percent on Terminal-Bench 2.1 and 78.5 percent on SWE-bench Multilingual, ahead of open models several times its size.

Start with the official demo Space on Hugging Face: paste in a gnarly function together with its failing test and watch it reason through the fix before answering. That gives you a feel for the model in two minutes with no account.

For real work, the model is on OpenRouter, and poolside's API speaks the standard formats, so most coding tools accept it as a drop-in backend. Quantized versions down to GGUF exist for running it yourself, but note the fine print: only 8B parameters are active per word, yet all 118B must sit in memory, so local means a serious workstation, not a laptop.

Most of the scores are poolside's own runs, and the closed frontier models still lead the hardest agentic tests by a wide margin, so this is the best value tier rather than the best absolute coder. And the small-model economics apply to speed and price, not memory: local use still needs workstation-class hardware.

• Drop a whole repository into the context and ask where a bug actually lives.
• Ask it to trace how a feature works end to end across dozens of files.
• Run long refactors in one session instead of re-briefing a model file by file.
• Swap it in as a cheaper engine behind the coding tool you already use.
• Run a quantized copy on your own hardware for code that cannot leave the building.

Try the demo

View model page

Read this issue on neuralexpedition.com →