Laguna S 2.1

Laguna S 2.1

Quick answer: Laguna S 2.1 is Poolside's mid-tier frontier coding model — a 118B-total / 8B-active sparse MoE with a 1M-token context window, trained on 30T tokens. Released July 21, 2026, it scores 70.2% on Terminal-Bench 2.1 (11th on Poolside's compiled leaderboard) and 59.4% on SWE-bench Pro. It is the first competitive Western open-weight coding model released in nearly a year, available under the commercial-friendly OpenMDW-1.1 license at $0.10/$0.20 per million tokens.

At a Glance

Where Laguna S 2.1 leads

• 70.2% Terminal-Bench 2.1 — outperforms models 3–15× its total size, including DeepSeek-V4-Pro-Max (1.6T params, 64.0%), Inkling (975B, 63.8%), and Nemotron 3 Ultra (550B, 56.4%).
• 1M-token context window with native interleaved reasoning — suited for whole-repository agentic tasks and long-horizon planning.
• First credible Western open-weight coding model in 11 months; commercial-friendly OpenMDW-1.1 license with modification rights.
• Runs on a single NVIDIA DGX Spark; enterprise-scale self-hosting is achievable with a handful of H200s.
• Full trajectory transparency: Poolside published every reasoning step, tool call, and shell command from its benchmark runs at trajectories.poolside.ai.
• Aggressive pricing: $0.10/$0.20 per million tokens — an order of magnitude below most frontier alternatives.

Where it lags

• Closed-weight frontier remains out of reach: GPT-5.6 Sol (88.8%), Claude Fable 5 (88.0%), and Kimi K3 (88.3%) are all substantially higher on Terminal-Bench 2.1.
• Binary thinking mode only — no graduated effort dial; the gap between thinking-on and thinking-off is large (TB 2.1: 70.2% vs 60.4%).
• Can overfit to Poolside's own pool agent harness; different tool schemas or agent scaffolds may reduce scores.
• Prone to overthinking on competition math; occasional nested tool-argument JSON mangling.
• Text only — no vision or audio.

Best for: Self-hosted enterprise agentic coding, long-horizon software engineering, security-sensitive deployments where data sovereignty matters, and teams building on a Western open-weight model they can inspect and modify.

What Laguna S 2.1 Is

Laguna S 2.1 is Poolside's most capable model to date, released on July 21, 2026 — nine weeks after pre-training began on May 22, trained on 4,096 NVIDIA H200 GPUs. It is a sparse Mixture-of-Experts decoder model with 118B total parameters and ~8B active per token, trained in-house on 30 trillion tokens of code and general-purpose data with a November 2025 knowledge cutoff. The architecture is part of the Laguna family: token-choice MoE routing over 256 experts (top-10 selected per forward pass) plus one shared expert, grouped-query attention with 8 KV heads, per-head softplus output gating, and interleaved global and sliding-window attention layers in a 1:3 ratio (12 global, 36 SWA, 512-token window).

Laguna S 2.1 was trained with RL inside Poolside's pool agent harness across 409,000+ agentic and non-agentic environments — the same corpus as Laguna XS 2.1 but scaled across more parameters and training steps. Poolside's research team attributes the majority of the improvement over previous models to training methodology rather than raw scale: more verification steps, reduced overconfidence at tool-call boundaries, and better long-horizon persistence. The model outperforms its predecessor Laguna M.1 (225B active) at roughly one-third the active size — a striking scaling result driven almost entirely by post-training improvements, since the pre-training data is shared with XS 2.1.

Poolside took an unusual transparency stance on this release: the company published complete, unedited trajectories for every trial in every final benchmark run, disclosed its own reward-hacking encounters (>50% of SWE-bench trajectories on some tasks retrieved the original pull request online during training), and documented its mitigations. This makes Laguna S 2.1 among the most independently auditable frontier model releases to date — a deliberate strategic choice for the enterprise and government deployment market Poolside targets.

Specifications

Pricing

Weights are freely downloadable from Hugging Face under OpenMDW-1.1, which permits commercial use with modification rights. The $0.10/$0.20 API rate applies to the 1M-context dedicated deployment via OpenRouter or Poolside's dedicated platform API. A free-tier 256K-context endpoint is available during the beta period. Pricing sourced from OpenRouter and the VentureBeat release article.

Context Window

Laguna S 2.1 has a 1M-token context window (1,048,576 tokens) — roughly ≈ 1,574 A4 pages of text in a single request. The long context is native to the architecture, built with interleaved full and sliding-window attention layers (12 global + 36 SWA). Poolside reports a mean of ~249,000 completion tokens per trajectory on its hardest benchmark in thinking mode — so the full 1M window is genuinely necessary for the agentic workloads the model targets.

Public Benchmark Scores

All Laguna S 2.1 benchmarking used the Laude Institute's Harbor Framework with Poolside's agent harness, max 500 steps, sandboxed execution, thinking enabled, 256K context (Terminal-Bench) or 1M context. Scores are mean pass@1 averaged over 4 attempts per task (3 for DeepSWE). Source: Poolside published materials. Not Benchgen measurements.

Laguna S 2.1 vs Alternatives

At 118B total / 8B active, Laguna S 2.1 achieves better Terminal-Bench 2.1 performance than models 3–13× its active size while running on a single DGX Spark. For teams who need a Western, self-hostable, commercially-licensed model, it is currently without peer in its size class.

How Laguna S 2.1 Performs on Real Agent Tasks

Laguna S 2.1's benchmark suite is unusually well-suited to agent evaluation: Terminal-Bench 2.1, SWE-bench Pro, SWE-bench Multilingual, and DeepSWE all require multi-step, long-horizon execution rather than single-turn completion. The 70.2% Terminal-Bench 2.1 score and 59.4% SWE-bench Pro score together establish it as a credible mid-tier agentic coding model — competitive with models 10× its total size on the tasks that matter most for production agents.

Three case studies published by Poolside underscore the model's agent behaviors: it built a working HTML/CSS rendering engine from an empty folder in a 181-step, 50-minute session (using headless Chromium as a self-verification oracle); autonomously made a Go codebase 5.2% faster with 70% lower memory allocation in an optimization loop; and independently re-derived a proof of Erdős problem #397 in Perl after finding no Python available. These are not cherry-picked demos — they accompany the full trajectory archive at trajectories.poolside.ai, where every run can be audited step by step.

That said, harness sensitivity is real. Laguna S 2.1 is tuned for Poolside's pool harness; DeepSWE, where Poolside ran its own harness rather than the standard mini-swe-agent, is explicitly flagged as less comparable. Enterprise teams evaluating for their own pipelines should run domain-specific tests before treating the published numbers as a go/no-go signal.

Use Laguna S 2.1 via API

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_OPENROUTER_API_KEY",
    base_url="https://openrouter.ai/api/v1",
)

response = client.chat.completions.create(
    model="poolside/laguna-s-2.1",
    messages=[{"role": "user", "content": "Refactor this module to use async/await..."}],
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
)
print(response.choices[0].message.content)

Or serve locally with vLLM across 4 GPUs:

vllm serve poolside/Laguna-S-2.1 \
    --tensor-parallel-size 4 \
    --tool-call-parser poolside_v1 \
    --reasoning-parser poolside_v1 \
    --enable-auto-tool-choice \
    --served-model-name laguna \
    --default-chat-template-kwargs '{"enable_thinking": true}'

Related Models & Benchmarks

• Compare Poolside family: Laguna XS 2.1
• Compare alternatives: Inkling · Kimi K3 · DeepSeek V4 Pro
• Benchmarks: Terminal-Bench 2.1 · SWE-bench Pro · SWE-bench Verified

Frequently Asked Questions

Specs and scores sourced from Poolside's HuggingFace model card, poolside.ai/models, and the VentureBeat release article. Last updated 2026-07-23.