Poolside has released Laguna S 2.1, an open-weight agentic coding model weighing 118 billion parameters. It operates as a Mixture-of-Experts system with 8 billion activated parameters per token and supports a context window of up to 1 million tokens. The weights are available on Hugging Face under an OpenMDW-1.1 license and fit within a single NVIDIA DGX Spark workstation.
On long-horizon coding benchmarks, the model performs alongside systems several times its size, including DeepSeek-V4-Pro-Max, NVIDIA’s Nemotron 3 Ultra, and Thinking Machines’ Inkling. Laguna S 2.1 extends the Laguna XS family, trained on the same pre-training data as the XS 2.1 variant.
Technical specifications
The model activates roughly 6.8% of its parameters for any given token. All 118 billion parameters remain resident in memory, but only approximately 8 billion route through the network per step. This sparsity allows a mid-size model to behave like a larger one while remaining cheap to serve.
Poolside publishes weights in BF16, FP8, INT4, and NVFP4 formats. Official GGUF and MLX conversions are included, alongside DFlash draft models. The team moved from the start of training to launch in under nine weeks. Pre-training began on 22 May 2026 across 4,096 NVIDIA H200 GPUs. This is the first Poolside model where reinforcement learning ran in FP8 precision.
Performance results
Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1 with thinking enabled. That places it first among open, disclosed-size models on Poolside’s leaderboard, behind only larger or closed systems. On SWE-Bench Multilingual it scores 78.5%, topping the published table outright.
| Benchmark | Laguna S 2.1 (118B-A8B) | Tencent Hy3 (295B-A21B) | Inkling (975B-A41B) | Nemotron 3 Ultra (550B-A55B) | DeepSeek-V4-Pro-Max (1.6T-A49B) | Kimi K3 (2.8T-A50B) | Qwen 3.7 Max | Muse Spark 1.1 | Claude Fable 5 |
|---|---|---|---|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 70.2 | 71.7 | 63.8 | 56.4 | 64.0 | 88.3 | 74.5 | 80.0 | 88.0 |
| SWE-Bench Multilingual | 78.5 | 75.8 | – | 67.7 | 76.2 | – | 78.3 | – | – |
| SWE-Bench Pro (Public) | 59.4 | 57.9 | 54.3 | – | 55.4 | – | 60.6 | 61.5 | 80.3 |
| DeepSWE v1.1 | 40.4 | – | – | – | 9.0 | 69.0 | – | 53.3 | 70.0 |
| SWE Atlas (Codebase QnA) | 46.2 | – | – | – | 27.2 | – | – | 42.2 | – |
| Toolathlon Verified | 49.7 | – | 45.5 | 34.3 | 55.9 | – | – | 75.6 | – |
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 are published at trajectories.poolside.ai.
Two thinking modes, and where the score comes from
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 reports coherent, productive reasoning running for hours and hundreds of thousands of tokens.
Three published runs
Poolside 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 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.
Deployment requirements
Sizing uses the full 118 billion parameters, not the 8 billion 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. 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, SGLang, and Ollama. Hosted access runs through OpenRouter, 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, Kilo, Prime Intellect Prime Lab, and ZML.
What it means
Developers can run the model locally on a single workstation rather than relying on a cloud API. The ability to activate specific experts per token means the system can handle complex code tasks without requiring massive hardware. However, the “max thinking” mode consumes significant tokens, so users must weigh the depth of reasoning against the cost per request.
