Poolside bets on Laguna S 2.1 for self-hosted coding AI

Poolside has launched Laguna S 2.1, a 118-billion-parameter open-weight model for agentic coding that the startup says can match or beat rivals several times larger while running on a single Nvidia DGX Spark desktop system.

The San Francisco company says the model uses a mixture-of-experts design with eight billion active parameters per token. Its weights are available on Hugging Face under the Linux Foundation’s OpenMDW license.

On two coding-focused agent benchmarks, Laguna S 2.1 posted results that put it in the same range as larger open models:

Recommended reading

Google Drops Gemini Sampling Controls in New Flash Models

• just over 70% on Terminal-Bench
• nearly 60% on SWE-Bench Pro

According to Poolside, that is enough to match or exceed models from DeepSeek, Nvidia, and Thinking Machines with two to eight times as many active parameters. The company also says the model is “not yet at the frontier,” with closed systems from OpenAI and Anthropic still scoring well above it on the same tests.

Western alternative to Chinese open-weight models

Poolside is framing the release as a direct answer to the dominance of Chinese labs in open-weight AI, especially DeepSeek, Alibaba’s Qwen family, and Moonshot’s Kimi. The company says no Western lab had released an open-weight model in the 118-billion-parameter class for 11 months before this launch.

Forbes reported that Poolside explicitly pitched the model as a self-hosted option for Western enterprises and governments that do not want to send data to a foreign provider. That fits the customer base Poolside is already targeting through its API and agent harness: government, defense, and other highly regulated organizations.

Founded in 2023 by former GitHub CTO Jason Warner and Eiso Kant, Poolside raised $500 million in a Series B in October 2024 at a $3 billion valuation, with backing from Nvidia and eBay. A planned $2 billion Series C that would have valued the company at $14 billion fell apart in April 2026 after CoreWeave exited a joint data center project in Texas.

Training on 4,000 H200 GPUs

Poolside says it built Laguna S 2.1 with its internal Model Factory platform, which automates architecture search and reinforcement learning from code execution. Training was completed in under four weeks on 4,000 Nvidia H200 GPUs.

The company released the smaller Laguna XS three weeks earlier and says it is shipping new models on roughly a five-week cadence. As a showcase for long-horizon reasoning, Poolside also published a trajectory showing the model independently solving a combinatorics problem that, until recently, only the largest frontier models had handled.

The commercial pitch is straightforward: customers may pay to run a capable coding model on their own hardware instead of relying on a closed API. But Poolside’s own benchmark data shows Laguna S 2.1 still trails closed-source leaders by roughly 10 to 15 percentage points on Terminal-Bench — a gap that could decide whether self-hosting is worth the trade-off.