Lux Automaton Logo
Lux Automaton
The AI Operating System
Laguna S 2.1: What a Long-Horizon Coding Model Can Actually Build
← All stories

Lux Automaton · AI Agent Architecture & Open Models

Laguna S 2.1: What a Long-Horizon Coding Model Can Actually Build

Poolside’s 118B open-weight model is designed to plan, use tools, recover, and keep working across large software tasks—if the hardware and agent harness are ready.
July 25, 20268 min read
Free activity or planner

Keep learning with the companion guide.

Open this simple resource in a new tab, then use it at your own pace.

Open the free guide →

The source video makes an exciting promise: build anything with Laguna S 2.1. That headline captures the direction of modern coding agents, but the practical version deserves more precision.

Laguna S 2.1 is an open-weight text-to-text model from Poolside designed for agentic coding, long-horizon software work, and extended tool-use workflows.

**Can this model complete one difficult job in my environment with acceptable quality, review time, hardware cost, and risk?**

01 — Long-horizon coding is more than generating a large answer

Long-Horizon Coding Loop
Long-Horizon Coding Loop

A long-horizon coding agent treats requests as a sequence of connected states: Inspect repo -> Find architecture -> Identify safe boundary -> Plan -> Edit -> Run tests -> Inspect result -> Recover -> Return evidence.

Reasoning between tool calls matters

Poolside's model card supports interleaved reasoning before and between tool calls. It does not blindly execute a fixed plan; it uses action results to decide the next step.

One million tokens is context—not guaranteed understanding

The base model lists a 1,048,576-token context window. A large context window allows holding deep code and docs, but tools must still direct model attention to the right files.

02 — '118B total, 8B active' changes the hardware conversation

Local Hardware & Model Reality
Local Hardware & Model Reality

Laguna S 2.1 uses a Mixture-of-Experts (MoE) architecture. The full model contains ~118B parameters, while ~8B are active per token.

Base BF16 weights require ~236 GB VRAM. Quantized INT4 and NVFP4 variants require ~71–72 GB before runtime overhead and KV cache.

03 — A production pilot matters more than a benchmark victory lap

Run a Production Pilot
Run a Production Pilot

Poolside reports 70.2% on Terminal-Bench 2.1 and 78.5% on SWE-bench Multilingual as of July 21, 2026. Benchmarks justify a pilot, not immediate production access.

Step-by-step pilot framework

1. Choose a bounded task: Upgrade a library, add one endpoint, or fix a reproducible bug. 2. Disposable workspace: Copy repo, test branch, separate credentials. 3. Acceptance criteria: Passing tests, max review time, rollback method. 4. Require evidence: Plan, diffs, commands executed, test results. 5. Keep production human-owned: A named human approves pull requests and deployments.

Laguna S 2.1 System Overview

Laguna S 2.1 Overview
Laguna S 2.1 Overview

Laguna S 2.1 brings serious long-horizon coding into open-weight models. Run a controlled pilot before adopting it as your default engine.

**“AI should work for you.” — Asa Pritchard** **Learn AI. Build Tomorrow. Change the World. — LANA**

Your next action

Open the Laguna S 2.1 Model Pilot Planner (/documents/laguna-s2-1-model-pilot-planner.html) and download the Model Pilot Template JSON (/documents/laguna-s2-1-model-pilot-template.json). Run one test repository task and compare Laguna S 2.1 with your current coding model.