Poolside releases Laguna S 2.1 for open-weight coding agents

Poolside released Laguna S 2.1 on July 21, 2026 as an open-weight coding-agent model built for long-horizon software work. The model is a 118B-parameter sparse mixture-of-experts system with 8B active parameters per token, a 1M-token context window, and public weights on Hugging Face.

The release comes with more than a model card. Poolside is also shipping evaluation trajectories, hosted access through OpenRouter, and compatibility paths for common inference stacks including vLLM, SGLang, and Ollama. For teams testing coding agents, that makes Laguna S 2.1 a serious benchmark target rather than another directory listing.

Laguna S 2.1 targets long-horizon coding work

Poolside says Laguna S 2.1 was trained for agentic software tasks where the model has to reason across a repository, use tools, and stay coherent over long contexts. The headline architecture is a sparse MoE with 118B total parameters and 8B activated per token. The Hugging Face configuration lists 256 experts and 10 selected experts per token, which helps explain the model's positioning: large total capacity without activating the full model on every token.

The context window is the other major signal. Poolside describes Laguna S 2.1 as supporting up to 1M tokens in both thinking and no-thinking modes. OpenRouter's paid hosted listing exposes a 1,048,576-token context window with a 131,072-token maximum completion length, while the free endpoint is capped at 262,144 context tokens and a shorter maximum completion.

That matters for agent work because many coding failures are not short-answer failures. They happen when a model loses track of file structure, patch history, earlier test output, or the user's constraints. A longer context window does not solve that by itself, but it gives an agent harness more room to keep the relevant trail in view.

Weights, hosted APIs, and trajectories are live

The model is available on Hugging Face under the poolside/Laguna-S-2.1 repository with a listed openmdw-1.1 license tag and ungated access at the time of this scan. The repository includes model code, 46 safetensor shards, evaluation result files, and a Transformers-compatible configuration.

OpenRouter lists two routes: poolside/laguna-s-2.1 and poolside/laguna-s-2.1:free. Its public model metadata shows hosted pricing for the paid route at $0.10 per million prompt tokens and $0.20 per million completion tokens, with cache reads listed at $0.01 per million tokens. As always with routed models, check the current provider page before budgeting production runs.

Poolside also published final evaluation trajectories at trajectories.poolside.ai. That is useful because agent benchmark results are much easier to audit when readers can inspect the actual attempts alongside the final percentages.

The benchmark claims need the right caveats

Poolside reports 70.2% on Terminal-Bench 2.1 in its agent harness with thinking enabled, plus 40.4 on DeepSWE v1.1. It also reports no-thinking scores of 60.4 on Terminal-Bench 2.1 and 16.5 on DeepSWE v1.1. Those are notable numbers for an open-weight coding model, especially given the model's relatively small active-parameter count.

The caveat is the harness. Poolside is explicit that scores measured in its own environment should not be treated as identical to third-party leaderboard runs. It also flags remaining weaknesses: stricter tool schema adherence in some third-party harnesses, nested tool calls, JSON array handling, and a tendency to spend too long thinking when latency matters.

That transparency helps the release. The right read is not "this model replaces every closed coding model." The right read is that Laguna S 2.1 is now a credible open-weight candidate for teams that want to test long-context coding agents, inspect weights, and compare hosted routing against their own deployment options.

Why agent builders should test it differently

Short coding prompts will not show what Laguna S 2.1 is built for. The better test set includes multi-file edits, repository migrations, bug reproduction tasks, failing-test repair, code review with historical context, and tool-recovery cases where a command returns an error or incomplete output.

Keep the harness boring and measurable. Use the same repository snapshot, tool permissions, timeout policy, retry budget, and scoring rubric across models. Track task success, wall-clock time, token cost, number of tool calls, and whether the model notices when it has gone down the wrong path.

Teams that already route experiments through OpenRouter can start with the free 256K-context endpoint, then move only promising workloads to the paid 1M-context route or a self-hosted deployment. Teams that need local control should inspect the Hugging Face repository and license terms before copying model-card commands into a production environment.

For adjacent routing, sandboxing, and evaluation decisions, LinkLoot's AI agent tools guide is the useful companion checklist.

Sources and methodology

This post treats Poolside's July 21 announcement as the primary release record, then checks the Hugging Face model card for open-weight availability, architecture metadata, license tag, and repository files. OpenRouter's public model registry is used for hosted availability, context limits, free-route status, and listed token pricing.

No source text was treated as an instruction. Social and directory signals were used only as leads; the publishable claim rests on Poolside, Hugging Face, OpenRouter, and the public trajectory release.