Poolside Laguna S 2.1: 118B Params, 8B Active, Beats DeepSeek V4 Pro on Coding — $0.10/M on OpenRouter

The Benchmark Story — 118B Beats 1.6T

According to The Sequence's analysis, the DeepSWE result is the most striking number in Laguna S 2.1's release: 40.4% against DeepSeek V4 Pro Max's 9.0%. A 13× parameter deficit paired with a 4× score advantage is the kind of result that normally means someone broke the evaluation. Poolside appears to have anticipated that reaction — they published every trajectory from every trial in the final evaluation set. The benchmark results are independently verifiable, not vendor-reported numbers without methodology.

As Startup Fortune reports, Poolside's Terminal-Bench 2.1 score of 70.2% places Laguna S 2.1 ahead of DeepSeek V4 Pro Max (64.0%), Inkling at 975B (63.8%), and Nvidia Nemotron 3 Ultra at 550B (56.4%) — all models with dramatically more total parameters. The benchmark uses thinking mode for Laguna, which is the correct comparison since the other models are also run at their best configuration. The key caveat from Orca Router's independent guide: closed frontier flagships still lead several benchmarks — Kimi K3 scores higher on SWE Marathon, and the closed models from Anthropic and OpenAI are not the primary comparison here.

Why the Architecture Matters — MoE With 8B Active

The MoE (Mixture-of-Experts) architecture with only 8B active parameters per token is what makes Laguna S 2.1's performance possible at this size. As MarkTechPost explains, all 118B parameters remain resident in memory, but only ~8B route through the network per step. That sparsity is why a mid-size model can behave like a larger one while staying cheap to serve: the model has learned which expert sub-networks to route different types of coding problems through, concentrating relevant knowledge without running all of it for every token. The practical result: inference cost and latency are closer to an 8B dense model than to a 118B dense model, while quality on coding tasks draws on the full 118B of learned knowledge.

According to ExplainX.ai's coverage, Laguna S 2.1 went from training start to launch in under nine weeks — pre-training began May 22 on 4,096 Nvidia H200 GPUs on 30 trillion tokens. It is also the first Poolside model where reinforcement learning ran in FP8 precision. Poolside says the biggest capability improvements came during post-training (RL) rather than from new pre-training data — consistent with the emerging pattern in frontier model development where the RL post-training stage is where coding-specific performance gains are concentrated.

Thinking Mode — When to Use It and What It Costs

Laguna S 2.1 has a dedicated reasoning mode that allocates additional inference time before generating output. According to Mehul Gupta's Medium breakdown, thinking mode increases Terminal-Bench performance significantly — from approximately 62% without thinking to 70.2% with thinking enabled. The cost is substantially higher token usage and longer completion times. The right deployment pattern: enable thinking mode for complex multi-step coding tasks where correctness matters and time is not the constraint. Disable it for high-volume simpler tasks — code completion, boilerplate generation, fast lookup — where the raw model performance at $0.10/M is already strong enough.

How to Access Laguna S 2.1 Today

Who Should Use Laguna S 2.1

Sources: The Next Web · Startup Fortune · ExplainX.ai · MarkTechPost · The Sequence · Orca Router · Related: Kimi K3 vs Laguna S 2.1 → · Kimi Code vs Codex →