Laguna S 2.1: How Poolside Built the Best 118B Agentic Coding Model
Poolside's Laguna S 2.1 is a 118B MoE model that scores 70.2% on Terminal-Bench 2.1 - beating models ten times its size. Here is what changed under the hood.
Laguna S 2.1: How Poolside Built the Best 118B Agentic Coding Model
Poolside released Laguna S 2.1 on 21 July 2026 - nine weeks after pre-training started on 4,096 NVIDIA H200 GPUs. The model is a 118B total-parameter Mixture-of-Experts (MoE) architecture with 8B activated parameters per token and a context window of up to 1M tokens. It supports both thinking and no-thinking inference modes.
The headline claim is straightforward: as of release, Laguna S 2.1 is the most capable agentic coding model in its weight class by a wide margin - and it holds that position against models that are 5 to 14 times larger in total parameter count.
Key Takeaways
- Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1 with thinking enabled, ranking 11th overall against all model sizes (Poolside, July 2026).
- Max thinking mode lifts Terminal-Bench performance from 60.4% to 70.2% and DeepSWE from 16.5% to 40.4%.
- Three models shipped in under three months using Poolside's internal Model Factory platform.
- Available open-weights on Hugging Face under OpenMDW-1.1 from day one of release.
1. How Does Laguna S 2.1 Perform Against Larger Models?
In July 2026, Laguna S 2.1 scored 70.2% on Terminal-Bench 2.1 in Poolside's agent harness with thinking enabled (Poolside, "Introducing Laguna S 2.1", July 2026). That places it 11th on the overall leaderboard - above models like Nemotron 3 Ultra at 56.4% (550B-A55B) and DeepSeek-V4-Pro-Max at 64.0% (1.6T-A49B). Both of those models have 4 to 13 times more total parameters.
The Terminal-Bench 2.1 benchmark evaluates long-horizon tasks where an agent model connects to its environment through a terminal. Tasks are designed to be hard to partially solve, which means scores actually spread and reveal meaningful differences between models.
On the harder DeepSWE v1.1 benchmark - where tasks are longer-horizon and harder to game - Laguna S 2.1 scores 40.4% in thinking mode. Frontier closed models GPT-5.6 Sol and Claude Fable 5 lead at 73% and 70% respectively. Laguna S 2.1 sits 13th, ahead of Gemini 3.5 Flash at 37% and DeepSeek-V4-Pro-Max at 9%.
The size-to-score scatter matters here. At 118B total parameters, S 2.1 outperforms a cluster of 200B-600B models. The compact size makes it feasible to run on a single NVIDIA DGX Spark workstation.
What does this mean for practitioners? A model you can run locally is categorically different from one that requires a datacenter. Laguna S 2.1 is the first sub-120B model that can credibly be used for production agentic coding work.
2. What Are the Real-World Task Results?
Poolside published three unedited case-study trajectories alongside the benchmark numbers.
2.1 Building a Browser Engine From Scratch
In a 50-minute session of 181 steps with no human intervention, Laguna S 2.1 built a working HTML/CSS rendering engine from an empty directory. The prompt asked the model to build the engine in vanilla JavaScript and prove it renders like a real browser by measuring against one.
The model built the full pipeline - parser, cascade, layout, and renderer - then found increasingly complex ways to validate its own output despite lacking vision capabilities. It ended up running headless Chromium to read canvas output back numerically for comparison. The model's approach was not prescribed; it arrived at that validation strategy on its own.
This case study illustrates a key property: Laguna S 2.1 finds clever paths to the goal even when the direct path is blocked.
2.2 Optimizing Poolside's Own Agent Harness
A Poolside researcher pointed the model at the internal agent harness used for training, evaluation, and user interaction. Over multiple hours in an automated research loop, Laguna S 2.1 made the harness 5.2% faster with approximately 70% lower memory allocation (Poolside, "Introducing Laguna S 2.1", July 2026).
The model identified that streaming-token accumulation used O(n^2) string concatenation and replaced it with buffers. It also found several instances of redundant copying and over-allocation during trajectory materialization, resolving them by memoizing materializations and pre-allocating slices to their exact sizes.
When wall-clock speedups became marginal and hard to measure, the model pivoted to memory allocation as a more accurately measurable target and kept improving. This is the persistence behavior Poolside explicitly trained for.
2.3 Re-Deriving an Erdos Problem Offline, in Perl
Laguna S 2.1 independently discovered a proof to Erdos problem number 397 (Erdos, Graham, Ruzsa, Straus, 1975), finding a construction that yields an infinite family of solutions. The sandbox had no Python installed. The model found Perl and conducted its number theory work there: brute-force exact prime factorizations, pattern analysis, a conjectured family, then the proof.
The result is a rediscovery - GPT-5.2 Pro published a solution in January 2026, the month after Laguna S 2.1's knowledge cutoff. Poolside notes that the construction S 2.1 derived is structurally different from the earlier published one: eight indices growing linearly versus the known six-index family, confirming a fresh derivation rather than recitation from training data.
3. How Does Thinking Mode Affect Performance?
Laguna S 2.1 ships with two inference modes: thinking off and thinking at maximum (enabled by default). The model determines its own thinking and test-time compute budget per problem.
The performance gap between modes is the largest Poolside has seen in any model it has trained to date:
- Terminal-Bench 2.1: thinking off scores 60.4%, thinking on scores 70.2% - a 9.8 point lift.
- DeepSWE v1.1: thinking off scores 16.5%, thinking on scores 40.4% - a 23.9 point lift.
Poolside has observed coherent, productive thinking over several hours and hundreds of thousands of tokens. The large delta on harder benchmarks suggests the internal monologue is most effective exactly where it matters: complex, long-horizon problems that require sustained reasoning.
User-configurable effort control (low / medium / high) is not included in this release. Poolside chose to ship without it to get the model into users' hands faster and plans to add it in a future iteration.
4. What Actually Changed in the Training Pipeline?
The improvements in S 2.1 come from behavioral training, not just scale.
Laguna S 2.1 is built on the same pre-training data as the XS 2.1 family. The step from XS 2.1 to S 2.1 involved scale, training-code fixes, and small recipe changes - not new pre-training data. S 2.1 is the first Poolside model where reinforcement learning was run in FP8 precision, which accelerated that stage.
Post-training ran in two stages. First, a supervised fine-tuning (SFT) stage bootstrapped capabilities using synthetic data. Second, RL was applied exclusively to tasks the model could not yet solve at a high pass rate.
The training corpus spans 409,000 agentic and non-agentic environments. Within this, 83,000 setups target terminal use cases and 168,000 target standard software engineering workflows. The largest single source reproduces real commits - approximately 38,000 tasks across 17,000 repositories.
Three changes in the training loop drove the persistence behavior:
- More generous rollout budgets: substantially longer timeouts, more tokens per turn, and more turns per task than any previous Poolside model.
- Better sandbox infrastructure: RL moved to a new sandboxing service with background processes, selective network blocking to reduce reward hacking surface, and artifact caching to avoid overloading external services.
- Multi-harness rollouts: the same prompts are rolled out in several agent harnesses, so the model learns behaviors that generalize across scaffolds rather than overfit to one.
5. What Are the Known Limitations?
Poolside published three known limitations with this release.
First, harness overfitting. In some cases S 2.1 struggles with adhering to tool schema definitions in third-party agent harnesses that differ slightly from Poolside's native harness. The model may rely on its memory of the tool interface on first use instead of following the provided definition. This is typically resolved through in-context learning if the harness rejects the invalid call and asks for a retry.
Second, nested tool calls. S 2.1 uses an XML-like tag format for tool calls. In cases where a tool argument expects a JSON array, the model may generate incorrectly escaped or invalid JSON.
Third, longer than expected thinking duration. S 2.1 may think for unusually long sequences before making progress, especially on competition mathematics problems. Future models will introduce effort control for thinking and improved thinking efficiency.
6. Where Can You Run Laguna S 2.1?
Laguna S 2.1 is available on Hugging Face under OpenMDW-1.1 with BF16, FP8, INT4, and NVFP4 weights, official GGUF and MLX conversions, and official DFlash draft models.
For open serving and local inference: vLLM, SGLang, and Ollama all support S 2.1 from day one. NVIDIA optimized inference across its hardware stack from TRT-LLM serving and NVFP4 on Blackwell systems down to a single NVIDIA DGX Spark.
For hosted access: Baseten's Model Library, OpenRouter (free endpoint at 256K context; paid at 1M context for $0.10 input / $0.20 output / $0.01 cache-read per 1M tokens), and Vercel AI Gateway all carry S 2.1.
For agentic coding tools: Kilo, Hermes Agent, pi, OpenCode, OpenClaw, Cline, and Poolside's own pool terminal agent all support S 2.1 from day one.
If you want the base pre-training weights for your own post-training, email models@poolside.ai.
7. Frequently Asked Questions
What size is Laguna S 2.1 and how many parameters are active per token?
Laguna S 2.1 has 118B total parameters in a Mixture-of-Experts architecture. Only 8B parameters are activated per token during inference, making it far cheaper to run than its total parameter count suggests. This design lets it run on a single NVIDIA DGX Spark workstation.
How does Laguna S 2.1 compare to models like DeepSeek-V4-Pro-Max in agentic benchmarks?
On Terminal-Bench 2.1, Laguna S 2.1 scores 70.2% versus DeepSeek-V4-Pro-Max at 64.0% - despite DeepSeek being 1.6T total parameters, more than 13 times larger. On DeepSWE v1.1, S 2.1 scores 40.4% versus DeepSeek at 9.0%, a 31.4-point gap (Poolside, July 2026).
What is Terminal-Bench 2.1 and why does it matter for evaluating agentic models?
Terminal-Bench 2.1 is a benchmark that evaluates wide, high-quality long-horizon tasks where an agent connects to its environment through a terminal. Tasks are hard to partially solve, so scores spread meaningfully across models. This makes it more diagnostic than benchmarks where top scores cluster tightly in the 70-90% range.
Is Laguna S 2.1 open weights?
Yes. It is available on Hugging Face under the OpenMDW-1.1 license with multiple quantization formats (BF16, FP8, INT4, NVFP4), official GGUF and MLX conversions, and official DFlash draft models. The base pre-training weights are available separately by emailing models@poolside.ai.
How long did it take Poolside to go from start of training to release?
Pre-training started on 22 May 2026 on 4,096 NVIDIA H200 GPUs. The model was released on 21 July 2026 - under nine weeks from start of training to public release.
References
- Poolside, "Introducing Laguna S 2.1", July 2026, https://poolside.ai/blog/introducing-laguna-s-2-1
- Poolside, "Laguna M.1 and XS.2 Technical Report", April 2026, https://poolside.ai/assets/laguna/laguna-m1-xs2-technical-report.pdf
- Poolside, Evaluation Trajectories, https://trajectories.poolside.ai
- Datacurve, DeepSWE v1.1 Leaderboard, https://deepswe.com
- Erdos, Graham, Ruzsa, Straus, 1975, Problem 397, https://www.erdosproblems.com/397
This article is based on the official Poolside release post for Laguna S 2.1, published 21 July 2026.
quizforml.com - Learn. Build. Fail. Learn Again.