Laguna S 2.1
Laguna S 2.1 is an open-weight Mixture-of-Experts model built for agentic coding and long-running tasks. It runs in thinking and no-thinking modes with a context window of 1M tokens. Call Laguna S 2.1 on AI Gateway with poolside/laguna-s-2.1. Your use is subject to Poolside's Terms & Privacy Policies.
import { streamText } from 'ai'
const result = streamText({ model: 'poolside/laguna-s-2.1', prompt: 'Why is the sky blue?'})Playground
Try out Laguna S 2.1 by Poolside. Usage is billed to your team at API rates. Free users (those who haven't made a payment) get $5 of credits every 30 days.
Laguna S 2.1
Laguna S 2.1
Providers
Route requests across multiple providers. Copy a provider slug to set your preference. Visit the docs for more info. Using a provider means you agree to their terms, listed under Legal.
P50 throughput on live AI Gateway traffic, in tokens per second (TPS). Visit the docs for more info.
P50 time to first token (TTFT) on live AI Gateway traffic, in milliseconds. View the docs for more info.
Direct request success rate on AI Gateway and per-provider. Visit the docs for more info.
About Laguna S 2.1
Laguna S 2.1 became available on AI Gateway on July 20, 2026. Laguna S 2.1 is a Mixture-of-Experts (MoE) model with 118 billion total parameters and 8 billion active per token, and it runs in thinking and no-thinking modes. Poolside published the weights on Hugging Face under the OpenMDW-1.1 license. AI Gateway lists two versions: this one, with a context window of 1M tokens, and a free version with a smaller window at laguna-s-2.1-free.
Agentic coding is the focus. With thinking enabled, Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, 78.5% on SWE-Bench Multilingual, and 59.4% on SWE-Bench Pro (public dataset). Broader agentic evaluations put it at 40.4% on DeepSWE v1.1, 46.2% on SWE Atlas codebase question answering, and 49.7% on Toolathlon Verified. Thinking accounts for a large share of those results. Without it, the Terminal-Bench 2.1 score falls to 60.4% and DeepSWE to 16.5%.
The context window of 1M tokens supports the sessions Laguna S 2.1 is built for. A single agent run accumulates terminal output, file contents, and tool results across many steps before it finishes. In one case study Poolside published, Laguna S 2.1 built a working HTML and CSS rendering engine from an empty folder over 181 steps with no human intervention, then checked its own rendering against a headless browser.
Poolside documented three limitations at release. First, Laguna S 2.1 sometimes relies on a remembered tool interface instead of the schema a third-party agent harness supplies. That usually corrects itself once the harness rejects the call and asks for a retry. Second, tool arguments that expect a JSON array can come back with invalid escaping. Third, Laguna S 2.1 can think for long stretches on competition mathematics before it makes progress. Laguna S 2.1 also accepts text input only, with no image or audio input.
Set the model to poolside/laguna-s-2.1 in the AI SDK, Chat Completions API, Responses API, Messages API, or other API formats, from TypeScript or Python. To use Laguna S 2.1 inside a coding agent, run vercel ai-gateway coding-agents setup, then select poolside/laguna-s-2.1 in the agent's model configuration. AI Gateway serves Laguna S 2.1 through Poolside, with retries and failover, and expanded capacity for both versions after launch. Pricing mirrors the provider with no markup and no platform fee on inference, including on Bring Your Own Key (BYOK) requests.
What To Consider When Choosing a Provider
- Configuration: Thinking mode is on by default, and Laguna S 2.1 chooses its own test-time compute budget for each problem. Poolside shipped this release without low, medium, and high effort levels, so the practical choice is thinking on or off. Hard problems can produce long thinking sequences, so set generous timeouts on agent runs and watch token usage in AI Gateway reporting.
- Zero Data Retention: AI Gateway does not currently support Zero Data Retention for this model. See the documentation for models that support ZDR.
- Authentication: AI Gateway authenticates requests using an API key or OIDC token. You do not need to manage provider credentials directly.
When to Use Laguna S 2.1
Best for
- Long-Horizon Coding Agents: Sessions that run for many steps and accumulate large working context
- Terminal-Driven Automation: Agents that work through a shell to build, run, and verify software
- Multilingual Bug Fixing: Repository work that spans several programming languages, with 78.5% on SWE-Bench Multilingual
- Codebase Question Answering: Questions that need the model to read widely across a repository first
- High-Volume Agent Traffic: AI Gateway expanded serving capacity for Laguna S 2.1 after the initial launch
Consider alternatives when
- Image or Audio Input: Laguna S 2.1 takes text only, so multimodal work needs a model such as Inkling
- Strict Third-Party Harnesses: Laguna S 2.1 can misread tool schemas that closely resemble the ones it knows
- Fixed Reasoning Budgets: Thinking is either on or off, with no low, medium, or high effort levels
- Broad General Assistance: Agentic coding is the target, not open-domain chat or factual recall
- Evaluation and Prototyping: The free version at
laguna-s-2.1-freecovers early testing at a smaller context window
Conclusion
Laguna S 2.1 is a compact agentic coding model that leans on thinking mode rather than parameter count. Use Laguna S 2.1 for terminal-driven agents, multi-language repository work, and sessions long enough to need a context window of 1M tokens. Try Laguna S 2.1 in the playground at https://ai-sdk.dev/playground/poolside:poolside/laguna-s-2.1:fast.