Poolside's Laguna XS 2.1 runs on a single Mac with 36 GB of RAM, scores 70.9% on SWE-bench Verified, and ships free on Hugging Face under a license that, for the first time, was actually designed for model weights rather than software code. That combination is worth paying attention to, even in a week where the open-weight leaderboard shuffles every few days.
The model dropped on July 2, 2026.
Poolside released it as a free download on Hugging Face and a free API tier on OpenRouter, the lowest-friction entry point the company has offered.
If you're currently running XS.2 through the Poolside or OpenRouter API, the upgrade path is clean:
pricing is matched, and the model ID changes from poolside/laguna-xs.2 to poolside/laguna-xs-2.1-no other integration changes required.
What actually changed from XS.2 to 2.1
Poolside kept the same underlying architecture: a 33-billion-parameter Mixture-of-Experts model that activates roughly 3 billion parameters per token, aimed at agentic coding and long-horizon work meant to run on a developer's own machine rather than a hosted cluster.
The headline number is on multilingual tasks. The 2.1 point release improves coding quality over XS.2: SWE-bench Multilingual moves from 57.7% to 63.1%, and SWE-bench Verified from 69.9% to 70.9%.
The multilingual gain-up 5.4 points-covers real GitHub issue resolution across Java, Go, Rust, TypeScript, and PHP repositories. Terminal-Bench 2.0 reaches 37.5%, up from 30.1% in the prior release.
Two architectural details make the local-inference story work. The model uses a mixed sliding-window and global attention layout in a 3:1 ratio across 40 total layers, with FP8 KV cache quantization to reduce memory per token, and native interleaved reasoning between tool calls that can be toggled per-request. Poolside also ships DFlash speculator models alongside the weights. These speculator models are open-weighted alongside each XS 2.1 checkpoint and reportedly double the achieved tokens-per-second for local inference. More specifically, the DFlash draft model proposes up to 15 tokens per step, with a reported end-to-end speedup of 1.67x-2.64x across evaluation datasets.
On benchmarking methodology: Poolside's benchmark claims come entirely from its own testing, run through the Laude Institute's Harbor Framework using its own agent harness, capped at 500 steps per task inside a sandboxed environment, with results averaged across multiple attempts per task.
It compared Laguna XS 2.1 against six models, including Qwen3.6, MAI-Code-1-Flash, gpt-oss-120b, Claude Haiku 4.5, and GPT-5.4 Nano-but most of those comparison scores came from each vendor's own release materials rather than an independent leaderboard. That is standard practice right now, but it means you should treat the relative comparisons with more skepticism than the absolute numbers.
Why the license change matters more than the benchmark bump
XS.2 shipped under Apache 2.0. XS 2.1 ships under OpenMDW-1.1, released by the Linux Foundation on May 28, 2026. This looks like a footnote. It is not.
Apache 2.0 was designed for software code, not AI model weights. Applying it to a set of trained parameters leaves several legal questions unanswered-whether patent rights are conveyed, whether database rights apply, what happens with model-generated outputs.
OpenMDW was built to close that gap. The OpenMDW license provides a single legal framework covering models (including architecture, weight, and parameters), code, documentation, and data.
Subject to its terms, OpenMDW-1.1 grants unrestricted, royalty-free permission to "deal in the Model Materials without restriction, including under all copyright, patent, database, and trade secret rights included or embodied therein."
It also explicitly addresses model-generated outputs-a critical gap in most existing open-source licenses.
NVIDIA simultaneously adopted OpenMDW-1.1 for its Cosmos, Isaac GR00T, Ising, and Nemotron model families , which is the real signal here. When the company that makes the hardware also backs the license, enterprise legal teams have cleaner ground to stand on. The Linux Foundation's own research estimates the global AI economy could save $20 billion to $48 billion a year if buyers simply chose the best model on price and performance-with licensing friction cited as one of the barriers.
The question for procurement teams is no longer just "can we run this?" It is "can we sign off on it?" OpenMDW makes the second question easier.
Deployment options and what fits where
Laguna XS 2.1 has launch-day support in vLLM, SGLang, Hugging Face Transformers, llama.cpp, and TRT-LLM.
At 33B total parameters and 3B activated, it is compact enough to run on a Mac with 36 GB of RAM and is available on Ollama and llama.cpp.
High-quality FP8, NVFP4, and INT4 quantized variants are available for teams with tighter VRAM budgets.
Paid API pricing is matched to XS.2: $0.10, $0.20, and $0.05 per million tokens for input, output, and cached input respectively, with 256K context length.
For teams starting fresh, Poolside recommends pool-its terminal-based coding agent-as the primary interface. The agent is open-source and ACP (Agent Client Protocol) compatible, the same harness Poolside uses internally for model training and evaluation. A teammate like Beagle that lives in Slack or Teams would surface pool's output in-channel without the developer needing to context-switch away from the conversation.
Laguna XS 2.1 activates only 3 billion of its 33 billion parameters per token, which Poolside positions against far larger dense systems like the 137-billion parameter MAI-Code-1-Flash, betting that sparse routing plus aggressive quantization wins local deployments where hosted-only systems cannot compete on cost or on-device control.
What is genuinely new versus incremental
The benchmark improvements are real but modest-a 5.4-point multilingual jump and incremental SWE-bench Verified gains are a point release, not a generational shift. The release lands as the open-weight coding tier increasingly competes on efficiency rather than raw size. In that context, the more durable change is the license. OpenMDW-1.1 is the first credible attempt to give enterprise legal teams a single document that covers weights, code, data, and outputs under one framework-and it has the Linux Foundation, NVIDIA, and now Poolside behind it.
Teams evaluating whether to bring a 33B coding model into a local workflow face a familiar decision tree: raw capability vs. inference cost vs. legal clarity vs. vendor lock-in risk. Laguna XS 2.1 does not top the raw capability charts- GLM-5.2, Z.ai's 744B MoE released in June, leads several long-horizon coding benchmarks and scores 91.2% on GPQA Diamond -but GLM-5.2 requires roughly 245 GB of combined memory at even its most quantized variant, and sits in a different procurement risk category for some teams. XS 2.1 fits on a MacBook Pro, works today in Ollama, and ships under a license a lawyer can actually read in five minutes.