Poolside's Laguna XS 2.1 Puts a Free Open-Weight Coding Model on Your Machine
Poolside released Laguna XS 2.1 on July 2, 2026 — a free, permissively licensed open-weight coding model that scores 70.9% on SWE-bench Verified and runs locally.
A Frontier-Class Coding Model You Can Actually Download
Every so often a release lands that quietly widens who gets to build with the best tools, and this is one of them. On July 2, 2026, Poolside published Laguna XS 2.1, a free open-weight coding model that any developer can download, inspect, and run on their own hardware. For those of us who care about accessible, private, capable AI, the details are genuinely exciting.
What Laguna XS 2.1 Is
Laguna XS 2.1 is a Mixture-of-Experts (MoE) model with 33 billion total parameters but only about 3 billion active at any moment, which is the architectural trick that lets it punch well above its running cost. It carries a generous 256K-token context window, so it can hold a large codebase in view while it reasons. The weights are available as a free download on Hugging Face and through OpenRouter, and crucially they ship under the OpenMDW-1.1 license — a new, fully permissive Linux Foundation license written specifically for model weights.
That licensing choice matters as much as the benchmarks. A permissive license means developers and companies can use the model in real products without restrictive strings attached, which is exactly the kind of openness that lets a broad community build on top of a release rather than merely admire it.
The Benchmark Numbers Are Serious
Capability is where an AI coding model earns its keep, and Laguna XS 2.1 shows up prepared. On SWE-bench Verified, the widely watched test of resolving real GitHub issues, it scores 70.9%. It reaches 63.1% on SWE-bench Multilingual — a 5.4-point jump over the previous version — 47.6% on the harder SWE-bench Pro, and 37.5% on Terminal-Bench 2.0. Those are numbers that would have been reserved for large proprietary systems not long ago, now available in a model small enough to run locally.
Built to Run on Your Own Terms
What I appreciate most is how deliberately the release is engineered for local AI. Laguna XS 2.1 runs across the popular local inference stacks — Ollama, llama.cpp, vLLM, SGLang, and TensorRT-LLM — in a range of precisions from BF16 down to INT4, so it scales from a workstation with a good GPU to leaner setups. Poolside also bundles DFlash speculator models that roughly double token generation speed through speculative decoding, a lossless technique that predicts several tokens at once and verifies them in a single pass.
Running a coding model locally is not just a performance preference; it is a privacy posture. Your proprietary code never leaves your machine, there is no per-token bill, and there is no dependency on an external service staying online. For solo developers, students, and teams with sensitive codebases, that combination of frontier-level capability and full local control is empowering.
Why This Release Matters
The through-line of 2026's most encouraging AI news is the shrinking gap between what the largest labs can do and what an individual can run at home. Laguna XS 2.1 is a clean example: strong agentic coding performance, a truly permissive license, and first-class support for the local tooling the open-source community already loves. When capable models are handed to everyone under open terms, the pace and diversity of what people build accelerates — and that is the version of AI progress worth celebrating.
Sources: Poolside Blog — "Introducing Laguna XS 2.1" — July 2, 2026; TechTimes — "Poolside Releases Free Open-Weight Coding Model" — July 4, 2026; Hugging Face — poolside/Laguna-XS-2.1 model card — July 2026.