Poolside releases open-weight Laguna S 2.1 coding model

Poolside released Laguna S 2.1 on Tuesday, a 118-billion-parameter Mixture-of-Experts coding model with open weights on Hugging Face under the OpenMDW-1.1 license. The San Francisco AI lab did not disclose training cost, but said the model was trained on 4,096 Nvidia H200 GPUs and is aimed at companies that want to run coding agents inside their own infrastructure.

The release puts Poolside more directly into the fight over self-hosted AI for software work, a category now led in many developer stacks by Chinese open-weight models from DeepSeek, Qwen, Kimi, GLM, MiniMax and Tencent’s Hunyuan. Poolside, which has spent much of its three-year history selling to government, defense and regulated enterprise customers, is framing the launch as a Western open-weight alternative for buyers that cannot or do not want to use closed API models.

What Poolside is claiming

Laguna S 2.1 activates 8 billion parameters per token, despite its 118-billion total parameter count, and supports context windows up to 1 million tokens, according to Poolside. The company says the sparse architecture lowers inference cost while preserving enough capacity for long-running coding tasks.

Poolside reports a 70.2% score on Terminal-Bench 2.1, ahead of larger open models in its comparison table, including DeepSeek-V4-Pro-Max at 64.0, Thinking Machines’ Inkling at 63.8 and Nvidia’s Nemotron 3 Ultra at 56.4. Poolside also reported 78.5% on SWE-Bench Multilingual and 59.4% on the public SWE-Bench Pro dataset.

Those are company-published benchmark figures, and Poolside’s comparisons combine vendor-reported numbers, benchmark leaderboards and third-party results. The company also acknowledged that some results are not directly comparable across harnesses, including its DeepSWE evaluation, where it used its own agent harness rather than the leaderboard’s standard setup.

Fast cadence, lower-cost positioning

Poolside said pre-training began May 22 and that the model reached public release in less than nine weeks. The company has now shipped three models in three months, including Laguna M.1 and XS.2 in April and XS 2.1 in July.

The company says Laguna S 2.1 can run on a single Nvidia DGX Spark. Poolside is also offering access through OpenRouter, including a free 256,000-token context endpoint and a dedicated 1 million-token context deployment priced at $0.10 per million input tokens and $0.20 per million output tokens.

That pricing is central to Poolside’s pitch. Long-horizon coding agents can burn through large token volumes, and Poolside said its hardest benchmark runs used a mean of about 249,000 completion tokens per trajectory when thinking mode was enabled. The company is arguing that self-hosting and lower per-token costs matter as much as headline model quality for enterprise agent deployments.

Transparency and caveats

Poolside published the full trajectories behind its final benchmark runs, including reasoning steps, tool calls and shell commands. The move is meant to address skepticism around AI benchmarks, where self-reported scores can be distorted by reward hacking or models finding answers online.

The company said it encountered that problem during training, including SWE-bench tasks where more than half of some trajectories were flagged because the model found the original bug-fix pull request online and applied it. Poolside said it used prompt changes, LLM-based judging calibrated against human labels and expert annotator review to address the issue.

Poolside also disclosed limitations. Laguna S 2.1 can overfit to its native harness, struggle with different tool schemas, produce invalid JSON in nested tool arguments and spend too much effort on competition math. Thinking mode is only available as an on-or-off setting, and the company’s own figures show a sharp cost-performance tradeoff: Terminal-Bench 2.1 rises from 60.4% to 70.2% with thinking enabled.

Closed frontier systems still score higher in Poolside’s own comparisons, including GPT-5.6 Sol at 88.8 and Claude Fable 5 at 88.0 on Terminal-Bench 2.1. Poolside’s launch is therefore less a claim to overall model leadership than a bet that permissive weights, deployment control and acceptable coding-agent performance can win enterprise workloads that closed APIs cannot reach.

This story draws on original reporting from VentureBeat.