Poolside Laguna S 2.1: Open-Weight Coding Model

Laguna S 2.1 specs and what the OpenMDW-1.1 license allows

Laguna S 2.1 is a coding model from Poolside whose weights are open to download. Open weights means the trained model files are published, so anyone can run the model on their own hardware instead of only through the maker's service. Poolside released it on 21 July 2026, and the weights were on Hugging Face from day one.

The model has 118B parameters in total but uses only 8B of them for each token. A token is a small piece of text, often part of a word. The context window, which is how much text the model can take in at once, goes up to 1M tokens.

The small active count comes from a mixture-of-experts design. The model holds many smaller expert networks and sends each token through only a few of them, and the model card lists 256 experts with 10 picked for each token.

Poolside published the weights in BF16, FP8, INT4 and NVFP4, with official GGUF and MLX conversions, so the same model comes in several file sizes. The release also came quickly: pre-training began on 22 May 2026 on 4,096 NVIDIA H200 GPUs, and launch followed in under nine weeks.

Laguna S 2.1 is a scaled-up version of Laguna XS 2.1, built on exactly the same pre-training data. It follows M.1 and XS.2, released together on 28 April 2026, and XS 2.1 on 2 July.

OpenMDW-1.1 is short, and its main terms fit in a few lines. It covers what it calls the Model Materials: the model itself, including its architecture and parameters, and related files such as data, documentation and software. The terms are these:

• Use: you may deal in the Model Materials free of charge and without restriction, including under copyright, patent, database and trade secret rights, as long as you follow the agreement.
• Redistribution: if you pass on any part of the materials, you must include a copy of the agreement and the copyright and origin notices that apply.
• Lawsuits: your rights end if you sue claiming that the materials infringe a patent or copyright, unless you are responding to a lawsuit that someone else brought first.
• Outputs: the license puts no restrictions or obligations on what the model generates.
• Warranty: the materials come "as is", and clearing other people's rights is the user's job.

In plain terms, a company can use the model, change it and build commercial products on it without paying Poolside or asking permission. The model card says the same in its own words.

The card also adds two points outside the license. Use should follow Poolside's Acceptable Use Policy. The model is designed for software engineering and agentic coding, meaning work where the model runs commands and edits files on its own, and the user must check that it suits their use.

Laguna S 2.1 benchmarks and how Poolside measured them

Laguna S 2.1 against DeepSeek-V4-Pro Max on all six benchmarks in the model card (%)

All six rows of the card's table, dated 21 July 2026. Laguna S 2.1 is higher on five; DeepSeek-V4-Pro Max leads on Toolathlon Verified.

70.2Laguna · Terminal-Bench 2.1

64.0DeepSeek · Terminal-Bench 2.1

78.5Laguna · SWE-bench Multilingual

76.2DeepSeek · SWE-bench Multilingual

59.4Laguna · SWE-Bench Pro

55.4DeepSeek · SWE-Bench Pro

40.4Laguna · DeepSWE

9.0DeepSeek · DeepSWE

46.2Laguna · SWE Atlas

27.2DeepSeek · SWE Atlas

49.7Laguna · Toolathlon Verified

55.9DeepSeek · Toolathlon Verified

0204060

A benchmark is a fixed set of tasks used to compare models, and each score here is the share of tasks solved. Poolside reports pass@1, which means the model gets one try per task, averaged over several runs. Most tests used four runs per task, and DeepSWE, SWE Atlas and Toolathlon Verified used three.

Every test runs the model inside a harness. A harness is the software that hands the model a coding task, lets it run commands and edit files, and then checks the result.

The model card prints all of the scores in one table, dated 21 July 2026. The version below keeps its six benchmarks and three of its comparison models. DeepSeek-V4-Pro Max is a much larger model, with 1.6T total parameters and 49B active, while Qwen 3.7 Max and Claude Fable 5 are listed without a size.

A dash means the model was not tested on that benchmark. A star marks a score that the card says came from a third party.

Read row by row, Laguna S 2.1 scores higher than DeepSeek-V4-Pro Max on five of the six benchmarks. The exception is Toolathlon Verified, where DeepSeek leads with 55.9 against 49.7 for Laguna.

On SWE-Bench Pro it sits below Qwen 3.7 Max (60.6) and well below Claude Fable 5 (80.3). The company's own ranking charts put it 11th of 22 models on Terminal-Bench 2.1 and 13th of 19 on the DeepSWE v1.1 chart.

Poolside's own claim is about models of its size, not about DeepSeek. It calls Laguna S 2.1, "as far as we can measure," the most capable agentic coding model in its weight class (models of a similar size) by a wide margin. The three rivals in the table above are either much larger or listed without a size, so the table does not test that claim directly.

The measurement details limit what the table can show. Laguna S 2.1's scores come from the company's own agent harness, called pool, with thinking on and a cap of 500 steps per task. Each run happened in a sandbox, an isolated machine where the model's commands cannot touch anything else.

For the other models, Poolside took the highest score it could find in the vendor's own report, the benchmark authors' leaderboard or the third-party site Artificial Analysis. The card marks four of DeepSeek-V4-Pro Max's six scores and one of Qwen 3.7 Max's as coming from third parties.

Poolside itself says this mix makes the scores less comparable. DeepSWE's leaderboard runs models in a harness called mini-swe-agent, while Laguna's 40.4% was scored in pool. The company adds that it does not believe this puts Laguna in a particularly favorable position.

Because each rival figure is the best one reported for it, the method does not pull rival scores down, and what differs is the harness. Poolside has also published the full trajectories, meaning the step-by-step record of every run behind each score, at trajectories.poolside.ai.

Max thinking raises scores and token use

Thinking means the model writes out its reasoning before it answers or acts. Laguna S 2.1 has only two settings, off and max, and max is on by default. There is no low, medium or high effort setting at launch; Poolside says it left that out to get the model to users right away and that later models will add effort control.

Max thinking accounts for a large part of the headline scores. It lifts Terminal-Bench 2.1 from 60.4% to 70.2%, and DeepSWE from 16.5% to 40.4%.

The cost is output length. On Poolside's chart, the average number of tokens the model writes per run grows from about 80k to 129k on Terminal-Bench 2.1, and from about 99k to 249k on DeepSWE.

Take a team that points the model at long bug-fixing tasks across a large codebase. Switching the reasoning off would cut the tokens it writes, but it would also drop the model to the lower no-thinking scores. For that kind of agent work, the model card recommends keeping thinking on.

The card also says to keep the model's earlier reasoning in the conversation history, because the model may stop reasoning in later steps if that history is dropped. The setting can be changed per request with enable_thinking.

Running Laguna S 2.1 on a DGX Spark or multiple GPUs

Quantization means storing each weight with fewer bits, which makes the model files smaller. BF16 is the full-precision release of Laguna S 2.1. FP8, INT4 and NVFP4 are lower-precision versions, and GGUF is the file format that llama.cpp reads.

Two statements about hardware need to be read side by side. The release post says the model is small enough to run on a single NVIDIA DGX Spark, and that NVIDIA helped tune inference all the way down to that machine.

The model card says the BF16 checkpoint needs multiple GPUs, at roughly 236GB of weights, and that quantized versions reduce this substantially. Its example vLLM command uses a tensor-parallel size of 4, which spreads the model across four GPUs. By the card's account, the full BF16 download is a multi-GPU job.

The card names five engines that can serve the model: vLLM, SGLang, Transformers, TRT-LLM and llama.cpp. For llama.cpp, it points to the company's own fork on the laguna branch and to GGUF files at poolside/Laguna-S-2.1-GGUF. Support in the main llama.cpp project is still under review (ggml-org/llama.cpp#25165).

Ollama users can start the model with ollama run laguna-s-2.1 and pick a quantized build by tag, such as q4_K_M, q8_0 or f16. Along with vLLM and SGLang, this route worked from day one, according to the release post.

Hosted access needs no hardware, because another company runs the model and you call it over the internet. At launch it was available through OpenRouter, Baseten's Model Library and Frontier Gateway, and Vercel AI Gateway.

The OpenRouter listing has a free endpoint with a 256K context window and a dedicated paid endpoint with the full 1M window. For a first look with no setup, chat.poolside.ai is a web chat with web search and basic code execution, and it needs no login.

A developer who wants to test the model on one repository this week does not need the 236GB download; the free OpenRouter endpoint or the web chat covers a first test. A team that must keep its code on its own machines needs a multi-GPU server for BF16, or one of the quantized builds.

Known limits: tool-call errors and overthinking

Poolside lists three known limits of Laguna S 2.1 in its own release post. Each one shows up in everyday agent work:

• Tool schemas in other harnesses: some third-party agent harnesses have tools that look like pool's but differ slightly, such as the terminal tool in Hermes Agent. There, the model may use the tool the way it remembers instead of following the definition it was given. This usually fixes itself if the harness rejects the bad call and asks the model to try again.
• Nested tool calls: the model marks tool calls with XML-like tags. When a tool argument must be a JSON array, as in Pi's edit tool, the model may write badly escaped or invalid JSON.
• Overthinking: the model may think for a long time before making progress, especially on competition math problems. Poolside says later models will add effort control and aim to think more efficiently.

Hermes Agent and pi also appear on Poolside's own list of tools that can run Laguna S 2.1 at launch. So the first two limits apply to setups on Poolside's own list, not only to unusual ones. Before wiring the model into one of these harnesses, run a few edit and terminal tasks and confirm that failed calls get retried.

Three things to check before you download the weights

The facts above come down to three checks, one for each decision a team makes before pulling the weights. Each check points back to the fact that settles it:

1. The license terms. If you will pass the model or a fine-tuned version to anyone, include the OpenMDW-1.1 text and the notices. Also check your use against Poolside's Acceptable Use Policy, which the model card points to. Outputs carry no license conditions.
2. How the scores were made. Treat the table as Poolside's own measurement: its own harness, max thinking on, and rival scores taken at their best reported value. On that table the model leads DeepSeek-V4-Pro Max on five of six tests and is not first on SWE-Bench Pro.
3. The route to running it. Plan for multiple GPUs if you want the BF16 weights. For a smaller setup, look at the quantized builds or Poolside's single-Spark claim. Leave thinking on for agent work, and expect the longer outputs that come with it.

Laguna S 2.1 is a 118B open-weight coding model whose best numbers come from Poolside's own harness at max thinking. Its license is short and permissive, and its full-precision build needs more than one GPU. Running open-weight models on your own hardware is covered in more depth in Local LLM Coding Now Hits 75% of Cloud AI.

The model card and the license text are ordinary web pages. To keep a clean Markdown copy of them next to your project notes, the tool below converts a page URL into Markdown.