Poolside released Laguna S 2.1 on July 21, 2026, a 118-billion-parameter Mixture-of-Experts model with roughly 8 billion activated parameters per token. The company claims it is the most capable agentic coding model in its weight class by a wide margin, and the benchmark numbers make a credible case.
On Terminal-Bench 2.1, which evaluates long-horizon tasks where an agent model is connected to its environment through a terminal, Laguna S 2.1 scores 70.2%. That places it ahead of DeepSeek-V4-Pro-Max at 64.0%, Nemotron 3 Ultra at 56.4%, and MiniMax M3 at 66.0%. It trails only frontier models like GPT-5.6 Sol at 88.8%, Kimi K3 at 88.3%, and Claude Fable 5 at 88.0%. On SWE-Bench Multilingual, it scores 78.5%. On SWE-Bench Pro, 59.4%. On DeepSWE, 40.4%.
The model went from the start of training to launch in under nine weeks, according to Poolside. It supports a context window of up to 1 million tokens in both thinking and no-thinking modes. The architecture uses 256 routed experts with a top-10 router plus one shared expert, 48 layers in a 1:3 ratio of global to sliding-window attention, and interleaved reasoning blocks emitted between tool calls.
The thinking trade-off
Laguna S 2.1 has two thinking modes. With thinking disabled, it scores 60.4% on Terminal-Bench 2.1 at an average of 80,000 completion tokens per trajectory. With thinking enabled, that jumps to 70.2% at 129,000 tokens. On DeepSWE, the gap is even wider: 16.5% without thinking at 99,000 tokens, versus 40.4% with thinking at 249,000 tokens. The model determines its own thinking budget rather than exposing user-configurable low-medium-high controls.
The licensing and deployment story
The weights ship under OpenMDW-1.1, which Poolside describes as permitting free commercial and non-commercial use and modification. That is a meaningfully more permissive posture than research-only or community licenses, and it is the specific unlock for enterprises that want an agentic coding model running inside their own network without a per-token bill.
The honest caveat is the hardware requirement. At 118B parameters, the BF16 checkpoint needs roughly 236GB of weights and multiple GPUs to serve. Open weights are not the same as cheap self-hosting for a small team. Poolside addresses this with quantized variants: FP8, NVFP4, INT4, and GGUF. DFlash draft models are also available to double achieved tokens per second.
The model is supported in vLLM, SGLang, NVIDIA TensorRT-LLM, and Ollama from day one. NVIDIA optimized inference across its hardware, from TRT-LLM serving and NVFP4 on Blackwell systems down to a single NVIDIA DGX Spark. For hosted access, it is available through Baseten's Model Library, OpenRouter, and Vercel AI Gateway. On OpenRouter, a free endpoint offers 256K context, while a paid endpoint provides the full 1M context window at 0.20 output per million tokens.
Three models in three months
Laguna S 2.1 is the third release in Poolside's rapid iteration cycle. Laguna M.1 and Laguna XS.2 shipped on April 28. Laguna XS 2.1, a 33B-parameter model, followed on July 2. Laguna S 2.1 arrived on July 21. Poolside says its internal Model Factory platform, a tightly integrated stack of versioned data, training, evaluation, and inference components, allows it to industrialize model development and push from one release to the next in weeks rather than months.
The company is explicit about its mission. It believes mastering agentic coding is the key to unlocking generalized autonomous problem solving for all knowledge tasks, since code is the most general medium for expressing and verifying solutions to structured problems. It also believes the West needs strong open-weight models, and it releases the Laguna family under permissive licenses from OpenMDW to Apache 2.0.
Coding benchmarks are leaky, and the model card does not publish long-context evals, which matters because the 1M window is one of the marquee specs. The forward-looking question is whether OpenMDW-1.1 becomes a template other labs adopt. If it does, the balance between hosted-only frontier coders and self-hostable ones shifts faster than the raw benchmark numbers alone would suggest.