Poolside drops open coding models Laguna XS.2 and M.1 alongside terminal agents

Laguna XS.2 is an open-weights coding foundation model from Poolside designed to run a fully functional coding agent on a single GPU.

With 33B total parameters and only 3B active during inference, XS.2 is released under the Apache 2.0 license on Hugging Face. It supports a 256K context window, making it accessible on single consumer-grade GPUs or local Apple Silicon setups without requiring multi-GPU server clusters.

Alongside XS.2, Poolside also launched Laguna M.1, a larger mixture-of-experts model pre-trained last year. M.1 packs 225B total parameters with 23B active, tailored specifically for long-horizon agentic software engineering and also served at 256K context length.

What it does and how to run it

Poolside paired the models with two developer products entering public preview.

The first is pool, a terminal-based coding agent. You can run Laguna XS.2 completely offline using Ollama, which features native MLX acceleration, turning your terminal into an interactive coding environment with a single launch command.

The second is Shimmer, an instant-on cloud sandbox preloaded with the Poolside Agent. It gives developers a turnkey virtual machine to test and iterate on web apps, APIs, and command-line utilities using both models.

If you prefer hosted endpoints, both Laguna M.1 and XS.2 are temporarily free via Poolside's developer platform API and on OpenRouter. Small research teams and startups can also request M.1 model weights directly from the lab.

Architecture and benchmarks

Laguna XS.2 was pre-trained from scratch inside Poolside's Model Factory in just five weeks before undergoing full post-training. The engineering behind it leans on the Muon optimizer, synthetic data automixing, and asynchronous on-policy agent reinforcement learning.

Both models were evaluated using the Harbor Framework across SWE-bench Verified, SWE-bench Multilingual, SWE-Bench Pro, and Terminal-Bench 2.0.

Benchmark comparisons put the lightweight XS.2 up against larger competitors like Qwen3.5 and Qwen3.6 35B models, Devstral Small 2, Gemma 4, Claude Haiku 4.5, and GPT-5.4 Nano. The heavyweight M.1 takes aim at frontier peers including GLM-4.7, DeepSeek-V4-Flash, Qwen3.5 397B, and Claude Sonnet 4.6.

Why it matters

For engineers and independent builders, Laguna XS.2 offers a rare sweet spot: genuine agentic coding capabilities that fit on a single GPU without routing proprietary source code through cloud APIs. Paired with Ollama and pool, it delivers an immediately usable local coding copilot.

For the broader ecosystem, open-weights releases in Western frontier labs have lagged behind closed services. By offering Apache 2.0 licensing and free hosted testing upfront, Poolside is pushing practical local agent workflows into daily development, though real-world stability across messy existing codebases still awaits wider community testing.