Laguna S 2.1: A Compact Agentic Coding Model That Punches Above Its Weight
Poolside today released Laguna S 2.1, a 118B total parameter Mixture-of-Experts (MoE) model with 8B activated parameters per token, supporting up to 1M token context in both thinking and no-thinking modes. The model went from training start to launch in under nine weeks. On long-horizon coding benchmarks, it holds its own against models many times its size.
Key Specifications and Benchmarks
Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1 in Poolside's agent harness with thinking enabled. This benchmark evaluates long-horizon tasks where an agent model interacts with its environment through a terminal. The model's compact size makes it suitable for complex work on local machines.
On DeepSWE v1.1, a benchmark with significant headroom where frontier models range from 54% to 73%, Laguna S 2.1 scores 40.4% in thinking mode using Poolside's harness. This is notable because some 1T+ parameter open models score below 10% on this benchmark.
| Benchmark | Laguna S 2.1 Score | Top Score |
|---|---|---|
| Terminal-Bench 2.1 | 70.2% | 88.8% (GPT-5.6 Sol) |
| DeepSWE v1.1 | 40.4% | 73.0% (GPT-5.6 Sol) |
| SWE-Bench Multilingual | (not disclosed) | - |
| SWE-Bench Pro (Public) | (not disclosed) | - |
All benchmark scores are pass@1 averaged over 4 attempts per task (except DeepSWE, SWE Atlas, Toolathlon Verified at 3 attempts). Full trajectories for every trial are available at trajectories.poolside.ai.
Evaluation Methodology and Transparency
Poolside has been vocal about reward hacking in benchmarks. For this release, they publish all trajectories from final evaluations of the published checkpoint, allowing independent verification. The evaluation methodology uses adversarial judging to increase reward hacking detection.
Case Study 1: Building a Browser Engine from Scratch
In a 50-minute session of 181 steps with no human intervention, Laguna S 2.1 built a working HTML/CSS rendering engine from an empty folder. It constructed an HTML tokenizer, DOM tree, CSS parser with selector specificity, cascade engine with inheritance, box-model layout, and canvas-2D renderer. The model then validated its work by running headless Chromium to read canvases and compare screenshots numerically. The final app renders nine HTML snippets side by side: the model's engine on canvas next to the hosting browser's iframe rendering for comparison.
Case Study 2: Optimizing Its Own Agent Harness
Poolside researchers pointed Laguna S 2.1 at their own agent harness (used for training/evaluation and user interaction). In an automated loop, the model made the harness 5.2% faster with ~70% lower memory allocation. It identified and replaced O(n²) string concatenation in streaming-token accumulation with buffers, and eliminated redundant copying via memoization and pre-allocation. The model continued optimizing even after speedup improvements became marginal, shifting focus to memory allocation when it became the more measurable metric.
Technical Details and Architecture
- Total parameters: 118B
- Activated parameters per token: 8B
- Context window: Up to 1M tokens
- Architecture: Mixture-of-Experts (MoE)
- Training time: Under nine weeks from start to launch
Comparison with Other Models
On Terminal-Bench 2.1, Laguna S 2.1 outperforms models like MiniMax M3 (66.0%), DeepSeek-V4-Pro-Max (64.0%), and Inkling (63.8%), despite being significantly smaller. It trails only the largest frontier models like GPT-5.6, Kimi K3, and Claude Fable 5.
On DeepSWE v1.1, the model's 40.4% score places it above Gemini 3.5 Flash (37.0%), Kimi K2.7 Code (31.0%), and Claude Sonnet 4.6 (30.0%), and far above DeepSeek-V4-Pro-Max (9.0%) and Laguna XS 2.1 (0.3%).
What This Means for Developers
Laguna S 2.1 offers a compelling option for developers who need a powerful agentic coding model that can run on local hardware. Its 8B activated parameters mean it can fit on consumer GPUs, while its 1M token context window allows it to work on large codebases. The model's ability to autonomously tackle long-horizon tasks—like building a browser engine or optimizing a harness—demonstrates practical utility for real-world engineering work.
Poolside has released all evaluation trajectories, enabling developers to inspect model behavior. The model is available as open weights (per the table), and developers can experiment with it on their own machines.
Getting Started
To try Laguna S 2.1, visit poolside.ai or check the trajectories at trajectories.poolside.ai. The model supports both thinking and no-thinking modes, and its compact size means it can run on a single GPU for many tasks.
Benchmarks as of 21 July 2026. All scores are pass@1 averaged over multiple attempts. For full methodology, see the original article.
