Poolside Releases Laguna XS 2.1

San Francisco startup Poolside announced Laguna XS 2.1 on July 2, 2026, an updated 33-billion-parameter Mixture-of-Experts model aimed at agentic coding and long-horizon tasks. The company published the release and model materials on its blog the same day.

Laguna XS 2.1 is described as a 33B total-parameter MoE with roughly 3B activated parameters per token, built with 40 layers and 256 experts and a very large context window. The model card lists a 262,144-token context and explicit architecture details intended for local and API use.

Poolside and the model card highlight a measurable performance bump versus the prior XS.2, most notably a 5.4-point jump on SWE-bench Multilingual to 63.1%. The Hugging Face model page shows the XS 2.1 table alongside XS.2 and other public models for comparison.

The release emphasizes on-device efficiency and several deployment-friendly features. Poolside and Hugging Face point to KV-cache quantized to FP8, launch-day support in vLLM and TRT-LLM, and multiple quantized checkpoints (FP8, INT4, NVFP4) to reduce VRAM and speed inference.

Poolside is making weights available on Hugging Face and serving the model through its API and OpenRouter, with free inference offered for a limited time to lower friction for developers. OpenRouter and Poolside documentation list the model slug poolside/laguna-xs-2.1 and instructions for invoking the model via an OpenAI-compatible API.

The company also changed the model license to OpenMDW-1.1 for XS 2.1, saying the move is intended to reduce licensing friction and enable broader reuse. Poolside framed the license switch as a step toward more permissive open-model distribution.

On the systems side, Poolside says it trained and evaluated Laguna models using a multi-stage pipeline that included pretraining, post-training, and reinforcement learning phases, and it points technical readers to a full technical report for details on data automixing and async off-policy agent RL. The report and model card are provided as supporting material.

For developers focused on local workflows, Poolside and Hugging Face claim XS 2.1 is compact enough to run on a Mac with 36 GB of RAM when using quantized checkpoints and the right runtime. The model card and blog also recommend Ollama, vLLM, and other local runtimes for best results.

Poolside disclosed the benchmarking method and several caveats: evaluations used Laude Institute’s Harbor Framework, patched task images to avoid infrastructure reliability issues, and ran tasks in sandboxed environments with fixed sampling parameters. The company notes it used highest publicly-referenced scores for comparison and flags where leaderboard data came from vendor posts.

Poolside plans a short transition for the older XS.2: the blog said Laguna XS.2 will sunset on the company API one week after the 2.1 launch, and coverage of the rollout repeated that July 9, 2026 retirement date for the older endpoint. Poolside said XS.2 will remain available in other libraries for dedicated deployments.

The immediate takeaway for engineers is practical: Poolside now offers an open-weight MoE tuned specifically for coding agents that can be run locally or via hosted endpoints, and it has added quantized checkpoints and guessing mechanisms Poolside calls DFlash speculators to speed local token throughput. The company invited developers to download the collection on Hugging Face, try OpenRouter, or use the Poolside API and share feedback on Discord.