Self-hosted coding agents
Run an open-weight coding model within infrastructure your team controls, subject to its license and acceptable-use terms.
Independent tool overview
Laguna S 2.1 at a glance
Laguna S 2.1 is a large open-weight mixture-of-experts model built for agentic coding and long-horizon software work. It combines roughly 8B active parameters per token with a 1M-token context window, but self-hosting still requires substantial storage, memory, and inference expertise.
Visit the official Laguna S 2.1 site ↗
- Developer
- Poolside
- Best for
- Self-hosted agentic coding
- Model size
- 118B total, about 8B active
- Context window
- 1,048,576 tokens
- License
- OpenMDW-1.1
- Primary modality
- Text-to-text
Overview
What Laguna S 2.1 is
Poolside positions Laguna S 2.1 as the middle model in its Laguna coding family: larger and more capable than Laguna XS 2.1, but smaller than Laguna M.1. Its main audience is engineering teams that want to run a coding model inside their own infrastructure or through a compatible inference provider.
The model has 118B total parameters with about 8B activated for each token. It supports tool-oriented, multi-step coding workflows, interleaved reasoning between tool calls, and a 1,048,576-token context window for large repositories and long agent trajectories.
The weights are available in BF16 and several quantized formats, including FP8, NVFP4, INT4, GGUF, and MLX variants. Poolside's main BF16 repository is about 235 GB, so the model is not a lightweight local download despite its relatively low active-parameter count.
Poolside releases Laguna S 2.1 under OpenMDW-1.1, which it describes as permissive for commercial and noncommercial use. Teams should still review the license, acceptable-use terms, security controls, and deployment dependencies before adopting it in production.
Use cases
Who Laguna S 2.1 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Large-repository work
Use the 1M-token context window for codebase exploration, cross-file changes, and long agent sessions.
Custom inference stacks
Deploy through supported runtimes such as vLLM, SGLang, TensorRT-LLM, llama.cpp, or Ollama.
Model evaluation teams
Test Poolside's model and quantized variants against your own repositories, tools, and software-engineering tasks.
Capabilities
Core Laguna S 2.1 features
Long-horizon coding
The model is post-trained for multi-step software tasks and agent workflows rather than simple code completion alone.
1M-token context
A 1,048,576-token context window can accommodate large codebases and extended tool-use histories, although usable performance still depends on the serving stack.
Mixture-of-experts architecture
Laguna S 2.1 has 118B total parameters while activating roughly 8B per token to balance capability and inference cost.
Controllable reasoning
Its chat template supports interleaved thinking around tool calls and per-request reasoning control.
Multiple weight formats
Poolside publishes BF16, FP8, NVFP4, INT4, GGUF, MLX, and speculative-decoding assets for different deployment targets.
Commercially usable weights
Poolside describes OpenMDW-1.1 as allowing use, modification, and commercial products without separate permission.
Process
How the Laguna S 2.1 workflow works
Step 1
Choose a checkpoint
Select BF16 for maximum fidelity or a supported quantized build that fits your available hardware and serving runtime.
Step 2
Deploy the model
Follow the official model-card instructions for a compatible inference engine and secure the endpoint before connecting tools.
Step 3
Connect a coding harness
Give the model tightly scoped repository, terminal, test, and search tools with explicit approval boundaries.
Step 4
Evaluate on your code
Measure correctness, latency, cost, security behavior, and long-context reliability on representative internal tasks.
Step 5
Add production controls
Keep human review, logs, secret filtering, sandboxing, and rollback paths around any agent that can modify or execute code.
Cost
Laguna S 2.1 pricing and free plan
Poolside publishes Laguna S 2.1's weights without a separate model download fee. Your real cost depends on storage, accelerators, inference software, operations, and any third-party hosting; Poolside does not list a simple public per-token price on the release or model card.
Open weights
No model download fee
Download the official checkpoints and run them in your own environment under OpenMDW-1.1.
- Infrastructure and operations are not included
- BF16 repository is about 235 GB
- Quantized variants can reduce storage and memory requirements
Self-hosted production
Infrastructure-dependent
Pay for the accelerators, storage, bandwidth, observability, and engineering needed to serve the model.
- Costs vary by hardware and traffic
- Long contexts materially increase compute and memory demand
- Benchmark your intended concurrency and latency
Hosted providers
Provider-specific
Compatible third-party inference services may offer Laguna S 2.1 with their own prices, limits, and retention policies.
- Confirm the exact checkpoint and context limit
- Review data handling before sending private code
- Compare total request cost, not only token rates
Pricing checked . Check current pricing at the source ↗
Assessment
Laguna S 2.1 strengths and limitations
Where it stands out
- Open weights give teams more deployment and data-location control than a closed API alone.
- The 1M-token context window is designed for large codebases and extended agent histories.
- Roughly 8B active parameters per token can make the 118B model more efficient than a dense model of the same total size.
- Official quantized builds and support for multiple inference runtimes broaden deployment options.
- Poolside publishes detailed architecture, deployment, responsible-use, and vendor benchmark information.
What to consider
- The full model remains large: the main BF16 repository is about 235 GB before runtime overhead.
- A 1M-token maximum does not guarantee accurate recall or efficient inference across every very long prompt.
- Poolside's benchmark results are useful directional evidence, but teams should reproduce performance with their own agent harness and repositories.
- Self-hosting transfers security, uptime, scaling, monitoring, and model-update work to your team.
- Coding agents can introduce vulnerabilities or destructive changes, so tool permissions and human review remain essential.
Compare
Laguna S 2.1 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Coding
Devstral
Consider Mistral's open coding-model family if you want another self-hostable agentic coding option.
Explore Devstral →Business Operations
DeepSeek
Consider DeepSeek for a broader open-model ecosystem with hosted API access and competitive coding capabilities.
Explore DeepSeek →Consumer
Claude Sonnet 5
Consider Claude Sonnet 5 if you prefer a managed frontier model and do not need to operate the weights yourself.
Explore Claude Sonnet 5 →Questions
Laguna S 2.1 FAQs
What is Laguna S 2.1?
Laguna S 2.1 is Poolside's 118B-total-parameter, open-weight mixture-of-experts model for agentic coding and long-horizon software work.
Can Laguna S 2.1 run locally?
Yes, if your workstation or server can handle the selected checkpoint. Quantized variants make local deployment more practical, but this is still a large model with significant storage and memory requirements.
How large is Laguna S 2.1's context window?
The official model card lists a 1,048,576-token context window. Actual speed, memory use, and task quality at long contexts depend on your serving setup and prompt.
Is Laguna S 2.1 free?
The official weights can be downloaded without a separate model fee under OpenMDW-1.1. Hardware, hosting, bandwidth, and engineering still create real costs.
Is Laguna S 2.1 open source?
It is best described precisely as an open-weight model under Poolside's OpenMDW-1.1 license. Poolside says the license permits commercial and noncommercial use and modification; review the license and acceptable-use terms for your case.
How should teams evaluate Laguna S 2.1?
Test it on representative repositories and agent tasks, then measure correctness, tool use, long-context reliability, latency, infrastructure cost, and security behavior before production use.
Bottom line
Our Laguna S 2.1 verdict
Laguna S 2.1 is a compelling option for teams that want a capable coding-agent model they can deploy and govern themselves. Its long context, quantized releases, and low active-parameter count are attractive, but the 118B checkpoint still demands serious infrastructure and careful agent security.
Visit Laguna S 2.1 website ↗Similar Coding Tools
View all Coding AI tools →
LM Studio Bionic
LM Bionic - LM Studio's local-first agent that codes and edits docs using open models
Cursor Router
Cursor Router - Cursor's auto-router picking the cheapest capable model per task
Canva Code
Canva Code 2.0 - Canva's vibe-coding tool with drag-and-drop editing
MAI-Cyber-1-Flash
MAI-Cyber-1-Flash - Microsoft's new cybersecurity model for securing codebases
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.