Poolside发布了Laguna S 2.1模型，这是一个118B总参数的混合专家模型，其中8B参数在推理时处于激活状态，支持100万token的上下文窗口 [1]。该模型从训练开始到发布仅耗时不到9周 [1]。
在编码基准测试中，Laguna S 2.1表现突出。在Terminal-Bench 2.1上的得分达到70.2%（启用思维模式），相比未启用思维模式的60.4%有明显提升 [1]；在DeepSWE v1.1上的得分为40.4%，较基础版本的16.5%大幅改进 [1]。这些成绩展现了该模型在编码任务上的竞争力，超越了多个体型更大的模型。
模型的训练数据基础广泛，共包含409k个代理和非代理环境，其中83k用于终端任务，168k用于软件工程工作流 [1]。软件工程任务中最大的数据来源是约38,000个真实提交任务，分布在17,000个代码仓库中 [1]。
除了编码能力外，Laguna S 2.1还展现了独立推理的能力。该模型独立重新推导了Erdős问题#397，并发现了无穷解族（在知识截断日期2025年11月之后）[1]。在自我优化方面，模型成功改进了自身代理框架的性能，速度提升了5.2%，内存分配降低了70% [1]。
在OpenRouter平台上，该模型的定价为输入token $0.10、输出token $0.20，缓存读取每100万token $0.01 [1]。
Poolside has unveiled Laguna S 2.1, a 118-billion-parameter mixture-of-experts language model designed to excel at extended reasoning and long-horizon tasks [1]. The model features 8 billion active parameters and supports a context window of 1 million tokens [1].
The development cycle proved remarkably efficient, with the model advancing from initial training to release in under 9 weeks, beginning on May 22 [1]. Despite its scale, Laguna S 2.1 demonstrates competitive performance on coding benchmarks compared to significantly larger models [1].
Performance and Capabilities
The model achieved a 70.2% score on Terminal-Bench 2.1 when operating in reasoning mode, up from 60.4% in standard operation [1]. On DeepSWE v1.1, it reached 40.4%, a substantial improvement from the baseline 16.5% [1]. These gains highlight the effectiveness of the model's reasoning capabilities [1].
The training dataset encompassed 409,000 agent and non-agent environment interactions, with 83,000 examples focused on terminal tasks and 168,000 on software engineering workflows [1]. Software engineering training drew from approximately 38,000 real commit tasks distributed across 17,000 repositories [1].
Noteworthy Achievements
Laguna S 2.1 independently re-derived Erdős problem #397 and discovered an infinite family of solutions, following a knowledge cutoff in November 2025 [1]. The model also successfully optimized its own agent framework, achieving a 5.2% speed improvement while reducing memory allocation by 70% [1].
On OpenRouter, pricing is set at $0.10 per million input tokens and $0.20 per million output tokens, with cached reads charged at $0.01 per million tokens [1].