Poolside 发布 Laguna S 2.1：118B 开源智能体编程模型，SWE-Bench Multilingual 表现超越同尺寸模型

Poolside：https://poolside.ai/ 发布了 Laguna S 2.1：https://poolside.ai/blog/introducing-laguna-s-2-1，这是一款拥有 1180 亿参数的开源权重模型，专为代理编码构建。它是一个专家混合（MoE）模型，每个 token 激活 80 亿参数。它支持在思考模式和非思考模式下最多 100 万个 token 的上下文窗口。权重可在 Hugging Face 上获取：https://huggingface.co/poolside/Laguna-S-2.1，采用 OpenMDW-1.1 许可证，模型体积足够小，可在单个 NVIDIA DGX Spark：https://www.nvidia.com/en-us/products/workstations/dgx-spark/ 上运行。

在长周期编码基准测试中，Laguna S 2.1 可以与几倍于自身规模的模型媲美，包括 DeepSeek-V4-Pro-Max、NVIDIA 的 Nemotron 3 Ultra 和 Thinking Machines 的 Inkling。Laguna S 2.1 是 Laguna XS 系列的扩展版，训练数据与 XS 2.1 相同。

该模型在任意给定 token 上大约激活 6.8% 的参数。全部 1180 亿参数都驻留在内存中，但每步仅约 80 亿参数通过网络路由。这种稀疏性使中等规模模型能够像大型模型一样运作，同时运行成本低。

Poolside 团队以 BF16、FP8、INT4 和 NVFP4 格式发布权重，同时提供官方 GGUF 和 MLX 转换以及 DFlash 草稿模型。从训练开始到发布仅用了不到九周的时间。预训练于 2026 年 5 月 22 日在 4096 个 NVIDIA H200 GPU 上开始。这是 Poolside 首个在 FP8 精度下进行强化学习的模型。

Laguna S 2.1 在启用思考模式的 Terminal-Bench 2.1：https://poolside.ai/blog/introducing-laguna-s-2-1 上得分为 70.2%。这使其在 Poolside 汇总的排行榜上位列开源、公开大小模型第一，仅次于更大或封闭系统。在 SWE-Bench 多语言测试中得分为 78.5%，直接排名公开表首位。Poolside 发布的完整对比如下。

最明显的信号来自 DeepSWE v1.1，它仍有提升空间。在该测试中，Laguna S 2.1 得分为 40.4%，而 DeepSeek-V4-Pro-Max 为 9.0%，活跃参数约为其六分之一。封闭前沿模型如 Claude Fable 5 和 Kimi K3 在多个基准测试中仍然领先。Poolside 所宣称的是权重类别上的表现，而不是绝对第一。最终评估集的轨迹发布在 trajectories.poolside.ai：https://trajectories.poolside.ai/。

Laguna S 2.1 有两种模式：关闭和最大，默认启用最大模式。在最大模式下，模型会自行设置测试时的计算预算。目前 Poolside 还没有提供用户可配置的低/中/高努力控制。

最大思考模式将 Terminal-Bench 2.1 从 60.4% 提升到 70.2%。它将 DeepSWE 从 16.5% 提升到 40.4%。这些提升需要消耗代币：DeepSWE 在思考模式下运行大约需要 249k 完成代币，而非思考模式下为 99k。Poolside 团队报告称，模型能够在数小时和几十万个代币内进行连贯、高效的推理。

Poolside 团队分享了三个未编辑的轨迹，以展示模型行为而非评分。在其中一个案例中，模型从一个空文件夹构建了一个可工作的 HTML/CSS 浏览器引擎。该运行在一次 50 分钟的会话中使用了 181 步，没有人工干预，并且模型将其输出在无头 Chromium 上进行了验证。

在第二个案例中，模型优化了 Poolside 自己的代理工具。它使工具速度提升了 5.2%，内存分配减少了约 71%，用缓冲区替代了 O(n²) 字符串拼接。在第三个案例中，它在 Perl 环境下离线重新推导了 Erdős 问题 #397：https://www.erdosproblems.com/397，因为沙箱环境没有 Python，耗时 68 分钟。该结果是独立的再发现；GPT-5.2 Pro 在 2026 年 1 月解决了同一问题，而 Laguna 的知识截止日期是 2025 年 11 月。

模型尺寸使用了完整的 118B 参数，而不是 8B 活跃参数，因为每个专家都保持在内存中。在 4-bit（NVFP4 或 INT4）下，权重大约需要 59 GB，这可以轻松放入单个 DGX Spark 的 128 GB 统一内存中。在 FP8 下需要大约 118 GB，仍可放入单个 Spark 或单个 H200：https://www.nvidia.com/en-us/data-center/h200/。在 BF16 下需要大约 236 GB，这就需要两个互联的 Spark 或多 GPU 数据中心节点。

Poolside 与 NVIDIA 合作，将 TRT-LLM 服务的推理优化到 Blackwell 上的 NVFP4，甚至只需单台 DGX Spark。它在第一天就支持了 vLLM：https://github.com/vllm-project/vllm、SGLang：https://github.com/sgl-project/sglang 和 Ollama：https://ollama.com/。托管访问通过 OpenRouter 运行：https://openrouter.ai/poolside，免费提供 256K 上下文，付费提供完整 1M 上下文，费用为每 1M 输入/输出/缓存读取令牌 $0.10/$0.20/$0.01。该模型也可以在 Baseten：https://www.baseten.co/、Kilo：https://kilocode.ai/、Prime Intellect：https://www.primeintellect.ai/ Prime Lab 和 ZML：https://zml.ai/ 上使用。

来源：Poolside 技术文章—推出 Laguna S 2.1：https://poolside.ai/blog/introducing-laguna-s-2-1，Robert McHardy 在 X 上：https://x.com/robert_mchardy/status/2079614440295506171，Poolside 在 X 上：https://x.com/poolsideai/status/2079613777343848465 以及 Hugging Face 模型卡：https://huggingface.co/poolside/Laguna-S-2.1。

Asif Razzaq 是 Marktechpost Media Inc. 的首席执行官。作为一位有远见的企业家和工程师，Asif 致力于利用人工智能的潜力造福社会。他最近的努力是推出一个人工智能媒体平台 Marktechpost，该平台以深入报道机器学习和深度学习新闻而闻名，既技术扎实又易于广大读者理解。该平台每月浏览量超过 200 万，显示了其受欢迎程度。

构建一个自主事件场地运营商 [完整代码]：https://pxllnk.co/twdn5

谢谢！我们的团队将很快与您联系 🙌

Poolside：https://poolside.ai/ has released Laguna S 2.1：https://poolside.ai/blog/introducing-laguna-s-2-1, a 118B-parameter open-weight model built for agentic coding. It is a Mixture-of-Experts (MoE) model with 8B activated parameters per token. It supports a context window of up to 1M tokens in both thinking and no-thinking modes. The weights are on Hugging Face：https://huggingface.co/poolside/Laguna-S-2.1 under an OpenMDW-1.1 license, and the model is small enough to run on a single NVIDIA DGX Spark：https://www.nvidia.com/en-us/products/workstations/dgx-spark/.

On long-horizon coding benchmarks, Laguna S 2.1 holds its own against models several times its size, including DeepSeek-V4-Pro-Max, NVIDIA’s Nemotron 3 Ultra, and Thinking Machines’ Inkling. Laguna S 2.1 is a scale-up of the Laguna XS family, trained on the same pre-training data as XS 2.1.

The model activates roughly 6.8% of its parameters on any given token. All 118B parameters remain resident in memory, but only ~8B route through the network per step. That sparsity is why a mid-size model can behave like a larger one while staying cheap to serve.

Poolside team publishes weights in BF16, FP8, INT4, and NVFP4, along with official GGUF and MLX conversions and DFlash draft models. It went from the start of training to launch in under nine weeks. Pre-training began on 22 May 2026 on 4,096 NVIDIA H200 GPUs. It is the first Poolside model where reinforcement learning ran in FP8 precision.

Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1：https://poolside.ai/blog/introducing-laguna-s-2-1 with thinking enabled. That places it first among open, disclosed-size models on Poolside’s compiled leaderboard, behind only larger or closed systems. On SWE-Bench Multilingual it scores 78.5%, topping the published table outright. The full comparison Poolside released is below.

The clearest signal is DeepSWE v1.1, which still has real headroom. There, Laguna S 2.1 scores 40.4% against DeepSeek-V4-Pro-Max’s 9.0%, with roughly one-sixth the active parameters. Closed frontier models such as Claude Fable 5 and Kimi K3 still lead on several benchmarks. Poolside’s claim is about the weight class, not the outright top. Trajectories from the final evaluation set is published at trajectories.poolside.ai：https://trajectories.poolside.ai/.

Laguna S 2.1 has two modes: off and max , with max enabled by default. In max mode the model sets its own test-time compute budget. Poolside is shipping without user-configurable low/medium/high effort control for now.

Max thinking lifts Terminal-Bench 2.1 from 60.4% to 70.2%. It lifts DeepSWE from 16.5% to 40.4%. Those gains cost tokens: DeepSWE trajectories run about 249k completion tokens in thinking mode against 99k without. Poolside team reports coherent, productive reasoning running for hours and hundreds of thousands of tokens.

Poolside team shared three unedited trajectories to show behavior rather than scores. In one, the model built a working HTML/CSS browser engine from an empty folder. That run took 181 steps across a 50-minute session, with no human intervention, and it validated its own output against headless Chromium.

In a second, the model optimized Poolside’s own agent harness. It made the harness 5.2% faster with roughly 71% lower memory allocation, replacing O(n²) string concatenation with buffers. In a third, it re-derived Erdős Problem #397：https://www.erdosproblems.com/397 offline in Perl over 68 minutes, since the sandbox had no Python. That result is an independent rediscovery; GPT-5.2 Pro solved the same problem in January 2026, and Laguna’s knowledge cutoff is November 2025.

Sizing uses the full 118B parameters, not the 8B active count, because every expert stays in memory. At 4-bit (NVFP4 or INT4) the weights need about 59 GB, which fits comfortably on a single DGX Spark’s 128 GB of unified memory. At FP8 they need about 118 GB, still within a single Spark or a single H200：https://www.nvidia.com/en-us/data-center/h200/. At BF16 they need about 236 GB, which calls for two linked Sparks or a multi-GPU datacenter node.

Poolside worked with NVIDIA to optimize inference from TRT-LLM serving to NVFP4 on Blackwell, down to a single DGX Spark. It shipped day-one support for vLLM：https://github.com/vllm-project/vllm, SGLang：https://github.com/sgl-project/sglang, and Ollama：https://ollama.com/. Hosted access runs through OpenRouter：https://openrouter.ai/poolside, free at 256K context and paid at the full 1M context for $0.10 / $0.20 / $0.01 per 1M input / output / cache-read tokens. The model is also on Baseten：https://www.baseten.co/, Kilo：https://kilocode.ai/, Prime Intellect：https://www.primeintellect.ai/ Prime Lab, and ZML：https://zml.ai/.

Sources: Poolside Technical— Introducing Laguna S 2.1：https://poolside.ai/blog/introducing-laguna-s-2-1, Robert McHardy on X：https://x.com/robert_mchardy/status/2079614440295506171, Poolside on X：https://x.com/poolsideai/status/2079613777343848465 and Hugging Face model card：https://huggingface.co/poolside/Laguna-S-2.1.

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.

Build an Agentic Event Venue Operator [Full Codes]：https://pxllnk.co/twdn5

Thanks! Our team will contact you soon 🙌