Laguna S 2.1 Release Notes: High-Performance Agentic Coding Model

Laguna S 2.1 is a 118B total parameter Mixture-of-Experts (MoE) model designed for long-horizon agentic work, featuring 8B activated parameters per token and a context window of up to 1 million tokens. It is positioned as the most capable agentic coding model in its weight class, competing with models significantly larger in total parameter count.

Benchmarking and Performance

Laguna S 2.1 demonstrates high efficiency in agentic coding tasks, often outperforming models with much larger parameter footprints.

Agentic Coding Benchmarks

On Terminal-Bench 2.1, which evaluates long-horizon tasks via a terminal interface, Laguna S 2.1 scored 70.2% with thinking enabled. This places it ahead of several larger open-weight models, including DeepSeek-V4-Pro-Max (1.6T parameters) and Inkling (975B parameters).

DeepSWE Performance

DeepSWE v1.1, designed to be more resistant to partial solutions and reward hacking, shows a wider spread of model capabilities. Laguna S 2.1 scored 40.4% in thinking mode using the "pool" agent harness. While trailing frontier closed models like GPT-5.6 Sol (73.0%) and Claude Fable 5 (70.0%), it significantly outperformed DeepSeek-V4-Pro-Max (9.0%) on this specific benchmark.

Summary of Key Benchmark Scores

Benchmark	Laguna S 2.1 (118B-A8B)	Tencent Hy3 (295B-A21B)	DeepSeek-V4-Pro-Max (1.6T-A49B)
Terminal-Bench 2.1	70.2%	71.7%	64.0%
SWE-Bench Multilingual	78.5%	75.8%	76.2%
SWE-Bench Pro (Public)	59.4%	57.9%	55.4%
DeepSWE	40.4%	-	9.0%
Toolathlon Verified	49.7%	-	55.9%

Core Capabilities and Case Studies

Laguna S 2.1 is characterized by its persistence and resourcefulness in solving complex, multi-step engineering problems.

Autonomous Software Engineering

In a documented case study, the model built a functional HTML/CSS rendering engine from a blank folder in a 50-minute session consisting of 181 steps. Without human intervention, it implemented a full pipeline (parser, cascade, layout, and renderer) in vanilla JavaScript and validated its work by comparing its output against a headless Chromium instance.

System Optimization

When tasked with optimizing its own agent harness, Laguna S 2.1 achieved a 5.2% speedup and a ~70% reduction in memory allocation. The model identified O(n^2) string concatenation in streaming-token accumulation and replaced it with buffers, while also resolving redundant copying during trajectory materialization.

Mathematical Discovery

The model independently re-derived a solution to Erdős problem #397, a conjecture that remained open for over 50 years. It utilized Perl for number theory and brute-force prime factorization to find a closed-form infinite family of eight-index solutions. This derivation was structurally different from the previously known six-index family, indicating a fresh discovery rather than memory recall.

Technical Architecture and Training

Laguna S 2.1 was developed in under nine weeks, scaling up from the Laguna XS family using the same pre-training data.

Post-Training and RL

The model's capabilities are primarily driven by a two-stage post-training process:

1. SFT Stage: Bootstraps capabilities using synthetic data.
2. RL Stage: Targets tasks with low initial pass rates. RL was performed in FP8 precision to accelerate training.

Training Data Distribution

The training corpus includes 409k agentic and non-agentic environments, with specific focuses on:

• Software Engineering: 168k tasks, including 38,000 tasks reproducing real commits across 17,000 repositories.
• Terminal Use Cases: 83k setups focused on building unseen environments.

Thinking Modes

Laguna S 2.1 features two modes: "off" and "max" (enabled by default). The "max" mode allows the model to determine its own test-time compute budget. Enabling thinking significantly boosts performance, raising the Terminal-Bench 2.1 score from 60.4% to 70.2% and the DeepSWE score from 16.5% to 40.4%.

Limitations and Known Issues

• Harness Overfitting: The model may struggle with tool schema definitions in third-party harnesses if they differ slightly from the native "pool" harness, occasionally relying on memory rather than the provided definition.
• Nested Tool Calls: The model may generate invalid or incorrectly escaped JSON when tool arguments require JSON arrays.
• Overthinking: In complex domains like competition mathematics, the model may enter long sequences of thinking before making progress.

Community Insights and Deployment

Community feedback highlights the model's accessibility for local hardware, with users noting its competitiveness against DeepSeek-V4-Flash and its ability to run on devices like the NVIDIA DGX Spark or high-RAM MacBooks.

"This is the first US release that's competitive with DeepSeek V4 Flash. Very excited about this."

Deployment Options:

• Open Weights: Available on Hugging Face under OpenMDW-1.1 (BF16, FP8, INT4, NVFP4).
• Local Inference: Supported by vLLM, SGLang, and Ollama.
• Hosted Access: Available via OpenRouter, Baseten, and Vercel AI Gateway.
• Direct Use: Accessible through the pool terminal-based agent or chat.poolside.ai.