Poolside Releases Laguna S 2.1, an Open Coding Model Aiming to Match Chinese Rivals

Poolside, a startup founded by former GitHub CTO Jason Warner, released the Laguna S 2.1 model under an open license on July 21. It's a Western lab's answer to Chinese coding models, which have been setting the pace in the open-weight segment for over a year.

Laguna S 2.1 is Poolside's third model released within three months, showing how much the company has accelerated its release cadence. The model targets agentic coding tasks, meaning it writes, tests and fixes code on its own without constant developer oversight.

Benchmark Results

On Terminal-Bench 2.1 with reasoning mode enabled, Laguna S 2.1 scores 70.2 percent, ahead of DeepSeek-V4-Pro-Max's 64.0 percent. On SWE-Bench Multilingual, the model reaches 78.5 percent, putting it in first place on a ranking published by Poolside itself, ahead of China's Tencent Hy3 (75.8 percent).

On SWE-Bench Pro, Laguna S 2.1 trails Qwen 3.7 Max by a narrow margin (59.4 versus 60.6 percent), but it clearly beats DeepSeek-V4-Pro-Max on the DeepSWE v1.1 and SWE Atlas tests. The authors openly acknowledge that the model isn't yet at the frontier, closed systems from OpenAI and Anthropic still score higher.

Turning on reasoning mode significantly boosts performance, but at the cost of more compute. On Terminal-Bench 2.1, the score rises from 60.4 to 70.2 percent, and on DeepSWE from 16.5 to 40.4 percent. Longer reasoning also means more tokens consumed, averaging around 249,000 in DeepSWE trajectories versus 99,000 without thinking mode.

Hardware and Availability

The model weights are available at several precisions. The 4-bit version takes up about 59 gigabytes of memory and fits on a single Nvidia DGX Spark machine, the FP8 version needs around 118 gigabytes, and full BF16 precision requires two linked Spark units or a data center node. The vLLM, SGLang and Ollama frameworks support the model from day one, with access also offered through OpenRouter, Baseten, Kilo, Prime Intellect's Prime Lab and ZML. Free access via OpenRouter includes a 256,000-token context window, while the paid tier offers the full million-token window priced at $0.10 per million input tokens and $0.20 per million output tokens.

The West needs open-weight models it can trust, run, and build on. Laguna S 2.1 is our answer - Jason Warner, co-CEO of Poolside

Laguna S 2.1 does the work of models several times its size because of how we build them, not in spite of it - Eiso Kant, co-founder and co-CEO of Poolside

Market Context

The launch comes after more than a year of Chinese labs dominating the open-weight category, with DeepSeek, Alibaba's Qwen and Moonshot AI's Kimi successively setting the pace. No Western lab had released a model in the 118-billion-parameter class in the previous 11 months, making Laguna S 2.1 the first response of its kind.

Poolside was founded in 2023 by Jason Warner, formerly GitHub's chief technology officer and head of engineering at Canonical and Heroku, and Eiso Kant, previously co-founder of several developer startups including Athenian. The company moved its headquarters to Paris after a $126 million seed round in 2023, and in October 2024 raised $500 million in a Series B round at a $3 billion valuation, with participation from Bain Capital Ventures, eBay and Nvidia. In October 2025, the company entered another round valued at roughly $12 billion, with Nvidia pledging an additional $1 billion in investment.

What It Means for Businesses and Developers

An open-weight model means companies can run it on their own infrastructure instead of sending queries to a closed API. That matters for public institutions and enterprises that, for security or regulatory compliance reasons, don't want to hand data over to foreign model providers, including Chinese labs.

For Polish tech companies and development teams, it offers another, cheaper alternative for running agentic coding tools locally, alongside earlier open Chinese models. The ability to run on a single DGX Spark device lowers the entry barrier for smaller companies without access to large GPU clusters.

Poolside also published three example sessions demonstrating the model's autonomous operation, including building a working HTML and CSS browser from scratch in 181 steps during a fifty-minute session, verified against headless Chromium, and speeding up Poolside's own agentic software by 5.2 percent while cutting memory usage by 71 percent.

The race for the best open coding model remains wide open. DeepSeek, Qwen and Kimi regularly update their models, and other Western labs, including Nvidia with its Nemotron series and Thinking Machines with its Inkling model, are also developing their own solutions in this segment. Poolside says further updates are coming as part of its accelerated release cycle, following three model launches in three months.