Poolside: Laguna S 2.1 (free)

Poolside Laguna S 2.1 is a mixture-of-experts language model designed for agentic coding and long-horizon work. It has 118B total parameters and about 8B active per token, placing it between the smaller Laguna XS 2.1 and larger Laguna M.1 in the manufacturer’s series. The model processes text and produces text, code, reasoning and tool-call requests in a tool-connected application.

Its architecture combines global and sliding-window attention across 48 layers: 12 global layers and 36 sliding-window layers. The 512-token sliding window describes local attention within those layers, while the documented model context is 1,048,576 tokens. These are different measurements. The larger context supports tasks that accumulate source files, observations and decisions over many steps. Keep the working objective and meaningful evidence together, and reserve space for subsequent reasoning and the final response within the applicable configuration.

The model’s central purpose is software work that requires persistence, checking and revision. For a repository task, provide the expected behavior, relevant files as text, dependency information and the actual failing output. Ask for a focused result: “Trace the cause of this regression, propose a change that preserves the documented interface, and identify the checks that would distinguish a fix from an unrelated change.” In a tool-connected workflow, give the model the result of each executed check before it chooses the next step.

For optimization work, provide measurements and an explicit objective. An example is: “Use this profile and allocation report to identify a bottleneck. Propose one change, explain the expected effect, and compare it with the baseline after the same measurement is repeated.” This makes a performance claim inspectable. Poolside’s release discusses verification and continued progress as important behaviors, but a persuasive explanation alone does not establish that a revision is faster, correct or complete. Preserve the observations that support keeping or rejecting the change.

Laguna S 2.1 supports interleaved thinking before and between tool calls. The manufacturer recommends preserving reasoning content from earlier assistant messages in subsequent history; dropping those blocks can cause reasoning to stop in follow-up steps. Keep the assistant response, tool-call information and the corresponding tool result together. This helps a longer task continue from its actual state rather than repeatedly reconstructing what was already attempted.

The documented thinking choices are off and max, with max enabled by default in the maker’s described setup. The model determines its thinking budget in that mode; the release does not present separate low, medium and high effort controls. For agentic coding, Poolside recommends enabling thinking and retaining it between steps. State the final artifact and completion criteria explicitly, and distinguish a long reasoning sequence from the completed result. The manufacturer identifies overthinking, especially on competition mathematics, as a substantive limitation.

Tool definitions need particular care. Poolside reports that S 2.1 can initially rely on a familiar tool interface instead of closely following a similar but different schema. It can also produce malformed or incorrectly escaped JSON when a tool argument contains a nested array. Give the actual argument definitions, validate proposed calls, and return a precise error when the call does not match. The maker describes correction through in-context feedback as a way to resolve some of these mismatches. Keep the resulting tool observation available for the next decision.

The manufacturer publishes separate BF16 and quantized variants and an optional DFlash draft model. Those are distinct artifacts and serving recipes, so a quantized checkpoint’s memory use or speculative-decoding performance should not be assumed from the general model name. For ordinary text work, prepare source code, command output and document extracts as readable text. The release notes the model’s lack of vision capability; a screenshot needs a textual observation or an appropriate external inspection step before it can support the analysis.

Request a final deliverable that identifies the changed artifact, the important decisions, the observed checks and any unresolved condition. For a code review, connect findings to the relevant source and show why each issue affects behavior. For a repair, separate the proposed patch from a patch whose behavior was actually checked. Retaining reasoning and tool history helps the model work through the task, while the final artifact and recorded results establish what was accomplished.