開放模型／程式代理
Poolside Releases Laguna S 2.1: 118B MoE Activates Just 8.5B Parameters per Token
Laguna S 2.1 uses sliding-window attention, attention gating, and an FP8 KV cache to reduce long-context costs, targeting self-hosted agentic coding. The model comes with weights at multiple precision levels and complete evaluation trajectories, but its million-token capability, performance comparisons, and single-machine feasibility still depend heavily on the quantization format and runtime environment.
Poolside has released Laguna S 2.1, an open-weight model designed for agentic coding. It has 117.6 billion parameters in total but uses a Mixture-of-Experts architecture that activates approximately 8.5 billion parameters per token. The model comprises 48 layers, 256 experts, and one shared expert, and supports preserving interleaved reasoning state between tool calls.
The architecture focuses not simply on scaling parameter count, but on controlling memory traffic for long contexts. Of its 48 layers, 36 use 512-token sliding-window attention, while only 12 perform global attention. A per-head softplus gate modulates the attention output, and the FP8 version can also quantize the KV cache to FP8. The official model card lists a native context window of 262,144 tokens, while some hosted endpoints offer up to 1 million tokens. Engineering teams therefore should not treat API specifications as guarantees that apply to every set of locally deployed weights.
Poolside reports that Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, 78.5% on SWE-bench Multilingual, and 59.4% on the public version of SWE-Bench Pro. It has also released the final evaluation trajectories, making it easier to verify whether an agent actually completed each task. However, the comparison table mixes results run by Poolside with figures reported by third parties, and the models, agent frameworks, time limits, and tool configurations are not fully consistent. Rankings therefore should not be based on a single percentage alone.
The weights are available in BF16, FP8, INT4, and NVFP4 formats, with support for vLLM, SGLang, TensorRT-LLM, llama.cpp, and Ollama. Even the FP8 files are approximately 121 GB. Claims that the model can “run locally” generally imply workstation-class unified memory, multiple GPUs, or further quantization—not a typical 24 GB graphics card. Its OpenMDW-1.1 license permits commercial use and modification, although deployers still need to review the usage restrictions. Key areas to watch next include independent long-horizon testing, capability loss after quantization, and the real-world cost of sliding-window attention when tracing dependencies across files in large codebases.