For Developers/Models/Laguna S 2.1
Open SourcePoolsideReleased 2026-07

Laguna S 2.1

The West's most capable open-weight model for agentic coding

Context window

1M

Input / 1M tokens

Not announced

Output / 1M tokens

Not announced

Provider

Poolside

Data verified 2026-07-22

A 118-billion-parameter open-weight Mixture-of-Experts (MoE) foundation model built for agentic coding. It activates 8 billion parameters per token and supports a 1 million token context window. The model matches or exceeds models several times its size on long-horizon coding benchmarks.

Capability index

Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.

Coding
9
Reasoning
7
Math
6
Multimodal
0
Long context
9
Speed
8
Cost efficiency
9

How to access it

Open-weight model available on Hugging Face under OpenMDW-1.1 license. Can be run locally on NVIDIA DGX Spark or accessed via OpenRouter.

Strengths

  • Exceptional performance on agentic coding benchmarks relative to model size
  • Efficient inference with only 8B active parameters despite 118B total
  • Supports 1M token context window
  • Small enough to run locally on single NVIDIA DGX Spark
  • Fully open-weight under permissive OpenMDW-1.1 license

Best for developers who...

Agentic coding tasksOrganizations requiring local deploymentCost-efficient coding model deploymentLong-horizon reasoning in coding scenarios

When to choose it (and when not to)

Reach for Laguna S 2.1 when...

  • Need open-weight coding model with strong benchmarks
  • Want to self-host rather than use API
  • Require 1M context window for coding tasks
  • Seeking competitive alternative to closed-source models

Look elsewhere if...

  • Require state-of-the-art frontier performance
  • Need general-purpose reasoning beyond coding
  • Can only use closed-source models

How to use it

  • Leverage thinking mode for complex multi-step coding problems
  • Utilize full 1M context window for long-context code understanding
  • Include agentic harness (pool) for autonomous coding task execution

Quickstart

Python
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained('poolside/laguna-s-2.1')
model = AutoModelForCausalLM.from_pretrained('poolside/laguna-s-2.1')

Weights also available in BF16, FP8, INT4, NVFP4 formats on Hugging Face

API model id: laguna-s-2-1

Benchmarks

BenchmarkScoreNotes
Terminal-Bench 2.170.2%With thinking mode enabled; 60.4% without thinking
SWE-Bench Multilingual78.5%Leads open disclosed-size models
SWE-Bench Pro~60%Matches or exceeds larger models
DeepSWE v1.140.4%With roughly 1/6 the active parameters of DeepSeek-V4-Pro-Max (9.0%)

Source: Poolside Official Announcement

Compare Laguna S 2.1

Compare Laguna S 2.1 with any other model

Build a comparison →
All model comparisons →

Learn the concepts