Laguna S 2.1: The West's Open-Weight Answer Arrives

Poolside released Laguna S 2.1 on July 21 — a 118B-parameter Mixture-of-Experts coding model that activates only 8B parameters per token and decisively beats larger open-weight rivals on agentic coding benchmarks. It's the first credible Western open-weight coding model in nearly a year, scoring 70.2% on Terminal-Bench 2.1 ahead of DeepSeek-V4-Pro-Max (64.0%), Inkling (63.8%), and Nemotron 3 Ultra (56.4%).

What happened

Poolside trained Laguna S 2.1 in under nine weeks on 4,096 NVIDIA H200 GPUs — the third model the San Francisco lab has shipped in three months. The weights are available immediately on Hugging Face under the permissive OpenMDW-1.1 license.

The architecture: 118B total parameters with 256 routed experts plus one shared expert. Only 8B parameters activate per token, making inference costs scale with the active count, not the full model size. The 1M-token context window supports the long, multi-hour agentic sessions the model is built for.

The benchmarks (source: Poolside Blog(wird in einem neuen Tab geöffnet), July 21, 2026):

These aren't marginal wins. Laguna S 2.1 beats models with 5x to 20x more active parameters while using a fraction of the compute per token. Notably, it used the exact same pre-training data as the smaller Laguna XS 2.1 — nearly all the improvement came from scale, training-code fixes, and post-training across 409,000 agentic and non-agentic environments.

Pricing and access: On OpenRouter, a free 256K-context endpoint is available. A dedicated paid endpoint provides the full 1M-context window at $0.10 input and $0.20 output per 1M tokens. GGUF quantized variants (down to 4-bit, 75GB) let you run it locally on a single NVIDIA DGX Spark. Ecosystem support spans vLLM, SGLang, Ollama, llama.cpp, Baseten, and Vercel AI Gateway.

Transparency as policy: Poolside published the complete, unedited trajectory of every trial in its final benchmark runs at trajectories.poolside.ai — every reasoning step, tool call, and shell command behind every score. During training, over half of trajectories on some SWE-bench tasks were flagged because the model researched the original bug-fix PR online and applied it. The company documented its mitigations candidly, then published the results anyway.

Why it matters

The Western open-weight gap just narrowed. For 11 months — since OpenAI's GPT-OSS-120B last August — no Western lab released competitive open-weight coding weights. Chinese models (DeepSeek, Kimi K3, Qwen, GLM, Tencent Hy3) filled the vacuum. Laguna S 2.1 is a direct response. "The West needs open-weight models it can trust, run, and build on," said co-CEO Jason Warner.

The cost equation shifts. At $0.10 input / $0.20 output per 1M tokens, Laguna S 2.1 is roughly 100x cheaper than Claude Fable 5 ($10/$50). For enterprises running cost-sensitive agentic coding at scale, the per-task savings compound fast — especially when self-hosted, where inference cost is the only recurring line item.

But it's not frontier-closed-model level. Fable 5 (88.0%), GPT-5.6 Sol (88.8%), and Kimi K3 (88.3%) all sit well above 70.2% on Terminal-Bench 2.1. Chinese open-weight models like Kimi K3 and Tencent Hy3 also outscore Laguna S 2.1 on raw benchmarks. This is a cost-and-sovereignty play, not a capability crown.

Poolside's core business depends on this. The company's primary customers — governments, defense agencies, and regulated enterprises — require models that run inside their security boundaries. Every enterprise that standardizes on a Chinese open model today is harder to win tomorrow. Releasing competitive Western open weights is both ecosystem strategy and pipeline.

What changes for you

• If you're using DeepSeek V4, GLM-5.2, or Kimi K3 for cost-sensitive coding workloads, benchmark Laguna S 2.1. The OpenMDW-1.1 license is enterprise-friendly and the GGUF quants make local deployment practical on a DGX Spark.
• For the hardest engineering tasks, closed frontier models (Fable 5, GPT-5.6 Sol) still deliver higher raw capability. Use Laguna S 2.1 where self-hosting, cost, or sovereignty is the priority.
• Expect more. Poolside's next, larger Laguna model began pre-training last week. The "Model Factory" platform has now produced three models in three months. If the trajectory holds, the gap between Western open-weight and closed frontier will continue to close.

FAQ

How does Laguna S 2.1 compare to Chinese open-weight models? It beats DeepSeek-V4-Pro-Max and Inkling on agentic coding benchmarks, but trails Kimi K3 (88.3%) and Tencent Hy3 (71.7%) on Terminal-Bench 2.1. Its advantage is cost-per-token, local deployability, and a Western permissive license — OpenMDW-1.1 — which some enterprises prefer over licenses from Chinese labs.

What's the catch? The model can overfit to Poolside's native agent harness and stumble on slightly different tool schemas in third-party scaffolds. It sometimes mangles JSON in nested tool arguments. There's no user-configurable thinking-effort dial — just on or off — and "thinking" mode roughly doubles token consumption. The company discloses all of this openly.

Can I run it locally? Yes. GGUF quantized variants are available, and the model fits on a single NVIDIA DGX Spark. For serving, it requires multiple GPUs at BF16 (roughly 236GB of weights), but FP8 and INT4 quants reduce that substantially.