Introducing Laguna S 2.1
- source_url: https://poolside.ai/blog/introducing-laguna-s-2-1
- source_type: product
- platform: blog
- author: Poolside
- original_date: 2026-07-21
- added_date: 2026-07-22
- category: models
- tags: laguna, coding-model, moe, long-horizon, poolside, trajectories
- quality_score: 4
- trajectories: https://trajectories.poolside.ai/
摘要（中文）
Poolside Laguna S 2.1：118B MoE、每 token 激活 8B、最长约 1M 上下文，训练到发布不到九周。强调长程 coding 与推理，并公开 final eval 全轨迹（trajectories.poolside.ai）。文中 Terminal-Bench 2.1 等分数与案例（含长步骤 canvas 引擎、自家 harness 加速）需注意 harness 边界，作者也提示 DeepSWE 等与 leaderboard harness 不可简单横比。价值在长任务验证循环与评测透明度，而不只是再报一个 coding 榜。
Summary (English)
Poolside releases Laguna S 2.1, a 118B MoE with 8B activated per token and up to 1M context, trained-to-launch in under nine weeks, aimed at longer-horizon work with reasoning. For published benchmark scores they release full final-eval trajectories. Terminal-Bench 2.1 and other numbers should be read with harness caveats the post itself notes (e.g. DeepSWE on pool harness vs leaderboard harness). Signal is long-task verification loops plus evaluation transparency, not a single leaderboard claim.
One-liner
Laguna S 2.1：118B-A8B 长程 coding，并公开 final eval 全轨迹。
Source body / metadata
Fetched via opencli web read / official docs during evening intake. Key claims are grounded in the source page and same-day Obsidian digest notes.
Poolside releases Laguna S 2.1, a 118B MoE with 8B activated per token and up to 1M context, trained-to-launch in under nine weeks, aimed at longer-horizon work with reasoning. For published benchmark scores they release full final-eval trajectories. Terminal-Bench 2.1 and other numbers should be read with harness caveats the post itself notes (e.g. DeepSWE on pool harness vs leaderboard harness). Signal is long-task verification loops plus evaluation transparency, not a single leaderboard claim.
Obsidian evidence
- local_note: OpenClaw定时任务/ClawFeed24小时高价值一览/2026-07-22-ClawFeed24小时高价值一览.md
- intake_run: daily-intake-evening 2026-07-22