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 coding tasks. By utilizing only 8B activated parameters per token, it achieves performance competitive with models many times its size, making it uniquely suitable for complex engineering work on local hardware.

Performance Benchmarks and Weight Class Comparison

Laguna S 2.1 is positioned as the most capable agentic coding model in its weight class. In evaluations using the Terminal-Bench 2.1 harness with thinking enabled, it scored 70.2%, outperforming significantly larger open-weight models such as DeepSeek-V4-Pro-Max (1.6T parameters, 64.0%) and Inkling (975B parameters, 63.8%).

Key Benchmark Results (pass@1)

Benchmark	Laguna S 2.1 (118B-A8B)	Tencent Hy3 (295B-A21B)	DeepSeek-V4-Pro Max (1.6T-A49B)	Claude Fable 5
Terminal-Bench 2.1	70.2%	71.7%	64.0%	88.0%
SWE-Bench Multilingual	78.5%	75.8%	76.2%	-
SWE-Bench Pro (Public)	59.4%	57.9%	55.4%	80.3%
DeepSWE	40.4%	-	9.0%	70.0%
Toolathlon Verified	49.7%	-	55.9%	-

On the DeepSWE v1.1 benchmark, which focuses on longer-horizon tasks that are harder to partially solve, Laguna S 2.1 scored 40.4% in thinking mode. This significantly exceeds the 9.0% score of the 1.6T parameter DeepSeek-V4-Pro-Max, demonstrating a superior ability to handle complex, multi-step software engineering problems.

Core Technical Specifications

Laguna S 2.1 was developed in under nine weeks from the start of training to launch. It is a scale-up of the Laguna XS family, utilizing the same pre-training data but incorporating updated training-code fixes and recipe changes.

Architecture: Mixture-of-Experts (MoE) with 118B total parameters and 8B active parameters per token.
Context Window: Supports up to 1M tokens in both thinking and no-thinking modes.
Training Precision: This is the first Poolside model where Reinforcement Learning (RL) was conducted in FP8 precision to accelerate training.
Post-Training: The model underwent a two-stage post-training process: an SFT stage bootstrapping capabilities with synthetic data, followed by RL for tasks with low initial pass rates.

Agentic Capabilities and Case Studies

Rather than focusing solely on raw intelligence, Poolside AI optimized for "behaviors that lead to a more capable model," specifically persistence, verification, and the willingness to backtrack.

Resourcefulness in Development

In one case study, Laguna S 2.1 built a functional HTML/CSS rendering engine from a blank folder in a 50-minute session (181 steps) without human intervention. Because the model lacks vision capabilities, it autonomously developed a validation pipeline using headless Chromium to compare its canvas renderings against a real browser numerically.

Engineering and Research Optimization

When tasked with optimizing its own agent harness, the model achieved a 5.2% speedup and a ~70% reduction in memory allocation. It identified O(n^2) string concatenation in streaming-token accumulation and replaced it with buffers, while also memoizing materializations to reduce redundant copying.

Mathematical Discovery

Laguna S 2.1 independently re-derived a proof for Erdős problem #397. Despite a knowledge cutoff of November 2025, it discovered a structurally different infinite family of solutions (eight-index) compared to the six-index family found by GPT-5.2 Pro in January 2026.

Thinking Modes and Test-Time Compute

Laguna S 2.1 features two thinking modes: Off and Max (enabled by default). The Max mode allows the model to determine the appropriate thinking/test-time compute budget for a given problem.

Thinking significantly impacts performance:

Terminal-Bench 2.1: Score increases from 60.4% (Off) to 70.2% (Max).
DeepSWE: Score increases from 16.5% (Off) to 40.4% (Max).

Known Limitations

Harness Overfitting: The model may occasionally rely on its memory of tool interfaces from its native harness rather than adhering to the specific tool schema definitions of third-party harnesses.
Nested Tool Calls: The model may generate invalid or incorrectly escaped JSON when a tool argument expects a JSON array.
Overthinking: The model may engage in excessively long thinking sequences before making progress, particularly with competition mathematics.

Community Feedback and Deployment

Community reactions on Hacker News highlight the model's efficiency and accessibility for local hardware. Users have noted its competitiveness with DeepSeek-V4-Flash and its ability to run on devices like the NVIDIA DGX Spark and Strix Halo.

"This is exactly the kind of model that's been needed in the middle. Realistically self-hosted, Good Enough intelligence, MoE so it's fast on limited bandwidth systems..."

Deployment Options:

Weights: Available on Hugging Face under OpenMDW-1.1 (BF16, FP8, INT4, NVFP4, GGUF, and MLX).
Hosted Access: Available via OpenRouter, Baseten, and Vercel AI Gateway.
Integrations: Supported by vLLM, SGLang, Ollama, and various coding agents including Cline, Hermes Agent, and pi.