Laguna S2.1: The American Open-Weight Model
Every time this series opens its doors to a new model, the disclaimer remains the same: this is not a scientific benchmark, but the report of a demanding user who runs an open-weight model on their home PC and puts it to the test with the same tasks reserved for past contenders. This time, however, the model's origin changes the weight of the question.
For months, those following the open model ecosystem have grown accustomed to a list that always sounds the same: Alibaba with Qwen, DeepSeek, Moonshot with Kimi, Zhipu with GLM. The center of gravity of open-weights—those downloadable, inspectable weights that can be run without asking anyone's permission—has shifted almost entirely toward Chinese laboratories. Poolside, a San Francisco company already known for serving enterprise and government clients, has decided to bring an American name back into play with Laguna S 2.1, a model with 118 billion total parameters and 8 billion active per token designed for agentic coding and long-horizon tasks, released on July 21, 2026, under the OpenMDW-1.1 license.
The tone Poolside adopted for the launch is not the subdued one of a simple technical update. Co-CEO Jason Warner framed the release as a response to a vacuum that the West built for itself: while the most capable US providers lock their best systems behind paid APIs and regulatory barriers, his company claims to offer a production-ready, open-weight option designed for those who need data control, predictable costs, and self-hosting capabilities. On the official Poolside blog, it is explicitly stated that the question of who provides the West's open models has moved from research circles to corporate boardrooms, and all the way to Washington. It is both an industrial and political framework: who controls the weights, who can run them on their own hardware, who can truly inspect them line by line.
My Setup, the Same Methodology
The hardware configuration remains the same as described in previous episodes of this series: a Ryzen 7700, 32 GB of DDR5 RAM, and a Radeon RX 9060 XT with 16 GB of VRAM—the same machine with which I have already tested Qwen 3.5, Qwen 3.6, the Gemma 4 family, and Ornith-1.0. Anyone looking for details on frameworks and selection criteria will find everything in the first episode and the second; here I will limit myself to confirming that the guiding principle remains unchanged: choosing the model size that can actually run on the machine and show its value in daily use, rather than the most impressive one on paper.
The Bigger Brother: Laguna S 2.1
Before diving into the size I actually tested, it is worth understanding where it fits within the family. Laguna S 2.1 has 48 layers, a declared context window of up to one million tokens, and shares its basic architecture with the rest of the family: a token-choice router with softplus gating over 256 experts plus one shared expert, grouped-query attention, and global attention layers alternated with sliding window layers. On the benchmarks declared by Poolside, Terminal-Bench 2.1 scores 70.2%, SWE-bench Multilingual 78.5%, and DeepSWE 40.4%—numbers which, according to MarkTechPost, place it at the top among open models of its declared size, though still behind closed frontier systems like GPT-5.6 Sol or Claude Fable 5, which remain above 88% on Terminal-Bench 2.1. In short, the Laguna family covers a range from the 33B XS size, designed to run locally, up to the 118B S size and beyond, with architectural consistency that makes the models comparable to one another.
The Test Model: Laguna XS 2.1
Laguna XS 2.1 is the compact size of the family: 33 billion total parameters with only 3 billion activated for each token, distributed over 40 total layers in a 3-to-1 ratio between sliding window attention (with a 512-token window) and global attention. Gating is sigmoid with rotational scales per layer, experts are 256 plus one always-active shared expert, and the KV cache is quantized in FP8 to reduce memory consumption per token. The declared context reaches 262,144 tokens, and the model is purely text-based—no vision modality—and supports native reasoning with thinking interleaved between tool calls, which can be enabled or disabled per individual request.
On official benchmarks, reported on the HuggingFace model card, Laguna XS 2.1 scores 63.1% on SWE-bench Multilingual, a jump of 5.4 percentage points compared to the previous XS.2 generation, and Poolside also claims a significant improvement on terminal-style tasks. For the sake of honesty with the reader, it must be clearly stated that some secondary sources circulating online report different figures for this model, including references to an architecture derived from Qwen2.5—information that primary sources from Poolside and HuggingFace explicitly deny, describing Laguna as a family with its own architectural recipe. I chose to stick to the verifiable data on the official pages rather than unconfirmed figures circulating elsewhere.
The model is distributed in BF16, FP8, INT4, and NVFP4, with declared support for vLLM, SGLang, NVIDIA TensorRT-LLM, HuggingFace Transformers, and Ollama, while native support for llama.cpp—and therefore for the GGUF files needed for LM Studio—arrived only later through a dedicated pull request on the official repository, as stated on the Poolside GGUF page. To complete the package, Poolside also released DFlash models, small speculators that, during local inference, double the actual tokens per second obtained in the company's tests.
OpenMDW-1.1: What "Open" Really Means
Laguna XS 2.1 is distributed under the OpenMDW-1.1 license, which stands for Open Model Data Weights—a license format backed by NVIDIA and the Linux Foundation that Poolside explicitly adopts to reduce legal friction in the distribution of open models, as explained on the official license page. This is a distinction worth making precisely because in journalistic language, "open source" and "open-weight" often end up confused. Here, there is no publication of training code or datasets; what is public and freely downloadable are the model weights, which can be used, modified, and redistributed even in commercial contexts without specific restrictions. It is the same principle already seen with licenses like Apache 2.0 or MIT for other models in this series, but applied to a format designed specifically for language model artifacts.
My Choice: GGUF Q6 on 16 GB of VRAM
The original BF16 version of Laguna XS 2.1 weighs 66.9 GB, a number that completely rules out any possibility of running it directly on my machine. The GGUF file table published by Poolside indicates Q4KM as the recommended default quantization for local use, about 20.3 GB, a compromise designed to remain accessible even on modest configurations, including Macs with 36 GB of unified RAM. For this test, I still chose a higher quantization, Q6_K, a step above the recommended default, to stay closer to the quality of the original model by leveraging the 16 GB of VRAM available on the Radeon. This was my own decision, not a preset indicated by Poolside; the official table does not list Q6 among the initially recommended options, but it is listed as compatible, and that is what I used in LM Studio, the same framework used throughout this series.
Eight Tests, a Halfway Verdict
Laguna XS 2.1 is Poolside's attempt to bring American open-weights back to the center of the debate, not with a generalist model, but with a compact, local system explicitly specialized in agentic coding. The eight tests that follow check in the field how well that declared specialization holds up against real use: coding, reasoning, long context, and multi-turn conversation quality.
LM Studio Configuration: context set to 150,784 tokens, GPU offload on 20 of the 32 layers, a pool of 8 CPU threads, evaluation batch of 2048, batch size 512, maximum of 4 concurrent predictions, 8 active experts per token.
Test 1, Scientific Reasoning on the Higgs Mechanism, Score 5/5, 15.26 tokens per second. The explanation unfolds in six logical sections, from the initial symmetry to the Higgs field with the "Mexican hat" potential, up to the mixing between the W3 and B fields and the definition of the Weinberg angle, both treated with precision. The only minor inaccuracy concerns the explanation for the photon's lack of mass, a detail that does not compromise the substance of an outstanding explanation for a model explicitly specialized in coding.
Test 2, Multimodal Reading of an Excel Table, Not Executed. Laguna XS 2.1 is a purely text-based model; loading the image in LM Studio is not recognized—a structural limitation consistent with Poolside's specifications, not a configuration defect.
Test 3, Code Generation for the Maximum Cycle in a Graph, Score 3.8/5, 11.60 tokens per second, with 13 minutes and 22 seconds of thinking. Here the model disappointed. The chain of reasoning shows a solid theoretical understanding: the NP-hard nature of the problem is recognized, as is the need to handle back-edges and track ancestors in the path to avoid duplicate counts. After thirteen minutes, however, the resulting algorithm solves a different problem than the one posed: it finds cycles in a DFS tree instead of the cycle of maximum length. A thinking time of this duration is in itself an anomaly not very compatible with interactive use, even more so when the final answer misses the mark.
Test 4, Multilingual Trip Planning, Score 3.5/5, 14.98 tokens per second. The prompt requested a response in French for a French customer, and here a curious dissociation emerged: the chain of thought plans correctly in French, but the final answer arrives in Italian on the first attempt. Only after an explicit prompt does the model correct the language, producing high-quality French with realistic references to Kinkaku-ji, Fushimi Inari, and the Tsukiji and Nishiki markets. Once the language is corrected, the content is solid, but the first-attempt behavior remains a warning sign regarding reliability in multilingual contexts.
Test 5, Long Context on a 460-Page PDF, Score 5/5, 13.88 tokens per second. The task was to load the Stanford HAI Artificial Intelligence Index Report 2025, approximately 460 pages, and ask the model for data on the growth of video generation, along with the page numbers where it is located. Despite a context set to 150,784 tokens, the model answered on the first attempt, pointing to pages 126 and 127—the same range identified by the best models tested so far in this series—citing Google Veo, Meta Movie Gen, and OpenAI Sora, the spaghetti eating test example, and even Figure 2.3.11, which compares user preferences between Veo 2, Movie Gen, Kling v1.5, and Sora Turbo. This is a level of detail that is above average.
Test 6, Spatial Reasoning on a Messy Room, Not Executed. Same limitation as Test 2, lack of multimodal support.
Test 7, Agentic Planning of a Web App, Score 4.8/5, 15.34 tokens per second. The proposed stack—React with Vite, Node.js with Express, PostgreSQL with Prisma, JWT, SendGrid, Puppeteer—is modern and coherent. The roadmap in four sprints is well balanced between setup, dashboard, PDF generation, and testing, with deliverables and critical issues indicated for each. Code examples for the Prisma model and the CSV import endpoint show real practical knowledge; the only missing piece is the explicit division of labor between the two developers requested by the prompt.
Test 8, Technical Conversation over Four Turns, Score 5/5, average speed in physiological decline from 15.54 to 11.45 tokens per second. Laguna maintained consistency by building a progressive architectural picture, from the initial stack to the comparison between WebSockets and polling for a thousand simultaneous users with code examples, down to a complete PostgreSQL schema with indexes and queries, closing with a scalability strategy to ten thousand users touching Node.js clustering, Redis sharding, load balancing, and monitoring.
Conclusions
The average score across the six completed tests stands at 4.52 out of 5, a respectable result that nevertheless reveals a model running at two different speeds. Where it needs to build, explain, plan, and maintain consistency over multiple turns, Laguna XS 2.1 behaves like a solid colleague, at times surprising on long contexts. However, where the task is the very one the model is sold for—pure coding on a medium-difficulty algorithmic problem—the performance fell short, with an execution time of thirteen minutes remaining the hardest detail to justify for anyone needing to use it in production.
In comparison with the other protagonists of this series—Qwen 3.5, the Gemma 4 family, Qwen 3.6, and especially Ornith-1.0, which closed with eight out of eight in this same series—Laguna XS 2.1 fails to establish itself as the benchmark for its size class, while remaining a model to keep an eye on precisely because of its origins. The most interesting question of all remains open: is the gap that still separates Western efforts from Chinese laboratories on open-weights a matter of time, with Poolside promising increasingly rapid release cycles thanks to its "Model Factory", or does it instead reflect a more structural gap in the amount of data and resources that Chinese companies can allocate to these projects without the constraint of having to sustain an enterprise business model in parallel?
Who should look at Laguna XS 2.1 today, with what hardware, and for which specific use cases is a question worth answering only after seeing how the next iteration behaves, considering that the previous generation, XS.2, arrived on the market only a few weeks before this one. Regarding the anomalous thinking time observed in Test 3, it remains to be understood whether this is a limitation of the model itself or a side effect of the Q6 quantization chosen for this test—a question that deserves dedicated verification before drawing final conclusions. And then there is the broader question, the geopolitical one from which we started: an American model capable of competing on local coding is big news, but is a single release enough to fill a gap that has formed over years of divergent strategies between the West and China on open artificial intelligence?