Poolside drops Laguna S 2.1: An 8B-active coding model punching way above its weight

Laguna S 2.1 is an open-weight coding model from Poolside designed to tackle grueling, multi-step engineering problems without breaking the bank on compute.

Under the hood, it is a 118B total parameter Mixture-of-Experts (MoE) model, but it only activates 8B parameters per token. That architecture keeps operational inference costs tiny while still handling an expansive 1M token context window—roughly 700,000 words of source code and technical context in a single prompt.

Poolside built and trained the model from scratch in under nine weeks. The weights are live right now on Hugging Face, and the team made the rare move of publishing full, unedited evaluation trajectories online so anyone can inspect its chain of reasoning.

What it can actually pull off

The standout trait of Laguna S 2.1 is persistence. It behaves less like an autocomplete engine and more like an autonomous developer that knows how to run terminal commands, inspect crashes, and course-correct on the fly.

The model ships with both thinking and no-thinking modes, with maximum thinking enabled by default. On Datacurve’s demanding DeepSWE benchmark, flipping thinking on propels its score from 16.5% all the way to 40.4%. On Terminal-Bench 2.1, it hits 70.2%, trading blows with proprietary models that pack vastly larger parameter counts.

Poolside shared several striking real-world traces. Given a blank directory and tasked with writing an HTML/CSS rendering engine in vanilla JavaScript, the model worked uninterrupted for 50 minutes and 181 steps. Lacking native vision, it spun up a headless Chromium instance to capture screenshots and mathematically compared canvas pixel dumps to confirm its CSS cascade worked properly.

In another test, researchers pointed it at their own agent evaluation loop. Over several hours of iterative profiling, the model spotted quadratic string concatenations, replaced them with pre-allocated buffers, and sliced memory allocations by roughly 70% while netting a 5.2% speedup. Even when trapped in a sandboxed offline shell without Python, it turned to Perl to independently construct a valid infinite family of solutions to Erdős problem #397.

Hardware specs and current catches

If you want to spin it up locally, the unquantized weights weigh in around 235 GB, demanding enterprise hardware like a DGX Spark. However, quantized builds are already landing, including an NVFP4 variant sitting at a much more manageable 71 GB.

While Poolside also rolled out a smaller 33B XS variant with 3B active parameters, early independent testing suggests it trails established local favorites like Qwen 3.6 35B. If you are experimenting with this architecture, the 118B model is where the real value lies.

Keep an eye on a few rough edges: the model can overfit to tool definitions it saw during training, occasionally fumbling tool schemas in third-party harnesses like Hermes Agent until kicked with an error message, and it can struggle with deeply nested JSON array inputs.

Why it matters

For engineering teams hesitant to feed proprietary codebases to hosted commercial APIs, Laguna S 2.1 offers a rare, viable path toward local autonomous coding agents. An 8B active footprint delivers real-time token throughput while the 1M context makes full-repo indexing feasible on private infrastructure.

We think Poolside made a very smart bet here. Instead of chasing multi-trillion parameter brute force, they optimized for raw agent endurance and tool proficiency at an accessible serving cost. It is rough around the edges on niche API schemas, but set loose in an automated test-and-repair loop, its problem-solving resilience is genuine.