Laguna S 2.1: Poolside's 118B Agentic Coding Model That Fits on a Single DGX Spark
AIOpen SourceLaguna S 2.1: Poolside's 118B Agentic Coding Model
On 21 July 2026, Poolside released Laguna S 2.1, a 118-billion-parameter Mixture-of-Experts model built specifically for agentic coding. With 8 billion active parameters per token and a 1-million-token context window, it fits on a single NVIDIA DGX Spark at 4-bit quantisation while competing with models orders of magnitude larger.
Architecture
Laguna S 2.1 is a scale-up of the Laguna XS family, trained on the same pre-training data as XS 2.1. It is a Mixture-of-Experts model — 8B active parameters out of 118B total — with support for both thinking and no-thinking inference modes. The thinking mode is enabled by default and sets its own test-time compute budget.
| Component | Specification |
|---|---|
| Total parameters | 118B |
| Active parameters per token | ~8B (MoE) |
| Context window | 1M tokens |
| Thinking mode | Enabled by default, automatic budget |
| Pre-training start | 22 May 2026 |
| Training duration | Under 9 weeks |
| Training hardware | 4,096 NVIDIA H200 GPUs |
| Precision (training) | FP8 (first Poolside model with FP8 RL) |
| Weight formats | BF16, FP8, INT4, NVFP4, GGUF, MLX |
The model is the first from Poolside where reinforcement learning ran entirely in FP8 precision, a sign that the industry's confidence in low-precision training has reached production maturity.
Benchmark Performance
Laguna S 2.1 punches well above its weight class. On SWE-Bench Multilingual, it scores 78.5% — the top of the published table among open, disclosed-size models. On Terminal-Bench 2.1, thinking mode lifts the score from 60.4% to 70.2%.
| Benchmark | Laguna S 2.1 (118B-A8B) | Kimi K3 (2.8T) | DeepSeek-V4-Pro-Max (1.6T) | Nemotron 3 Ultra (550B) |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 70.2 | 88.3 | 64.0 | 56.4 |
| SWE-Bench Multilingual | 78.5 | — | 76.2 | 67.7 |
| DeepSWE v1.1 | 40.4 | 69.0 | 9.0 | — |
| Toolathlon Verified | 49.7 | — | 55.9 | 34.3 |
The DeepSWE result is particularly informative. Laguna S 2.1 scores 40.4% against DeepSeek-V4-Pro-Max's 9.0% — with roughly one-sixth the active parameters. Closed frontier models such as Claude Fable 5 and Kimi K3 still lead on absolute scores, but the efficiency gap has narrowed dramatically.
Thinking Mode in Practice
Poolside shared three unedited trajectories that reveal how the model reasons:
-
Browser engine from scratch: The model built a working HTML/CSS browser engine from an empty folder. It took 181 steps across a 50-minute session with no human intervention, validating its own output against headless Chromium.
-
Agent harness optimisation: The model optimised Poolside's own agent harness, making it 5.2% faster with roughly 71% lower memory allocation by replacing O(n²) string concatenation with buffers.
-
Independent rediscovery: It re-derived Erdős Problem #397 offline in Perl over 68 minutes — the sandbox had no Python. GPT-5.2 Pro solved the same problem in January 2026; Laguna's knowledge cutoff is November 2025.
These trajectories demonstrate sustained, coherent reasoning over hundreds of thousands of tokens. DeepSWE trajectories average approximately 249k completion tokens in thinking mode against 99k without — a 2.5× compute investment that yields a 24-percentage-point gain.
Deployment on Consumer Hardware
The most significant operational detail for enterprise teams: at NVFP4 or INT4 quantisation, the weights require approximately 59 GB, fitting comfortably on a single DGX Spark's 128 GB of unified memory. At FP8, 118 GB still fits within one Spark or one H200. BF16 requires approximately 236 GB, which calls for two linked Sparks or a multi-GPU datacenter node.
Day-one support is available in vLLM, SGLang, Ollama, TRT-LLM, and NVIDIA's NVFP4 stack on Blackwell. Hosted access runs through OpenRouter at 0.20/$0.01 per 1M input/output/cache-read tokens.
Strategic Position
Laguna S 2.1 occupies a unique position in the 2026 model landscape: an open-weight coding agent that fits on a single workstation GPU while delivering scores competitive with datacenter-scale models. For enterprise teams building agentic coding pipelines, it represents the first viable on-premise alternative to API-based coding agents — without sacrificing capability.