Laguna S 2.1

Laguna S 2.1 is Poolside's open-weight coding model, released July 21, 2026 with weights on Hugging Face the same day. It pairs a 118B-total / 8B-active Mixture-of-Experts architecture with a 1M-token context window and native interleaved reasoning, targeting agentic coding and long-horizon work. Under the OpenMDW-1.1 license it is free for commercial and non-commercial use, and it sits in the Laguna family between Laguna XS 2.1 (33B-A3B) and Laguna M.1 (225B-A23B).

Model Specifications

Key Features

• Native interleaved reasoning: Thinking is woven between tool calls and controlled per request with enable_thinking. Reasoning runs either off or max; there are no low, medium, or high tiers.
• 1M context everywhere: The full 1,048,576-token window is available in both thinking and no-thinking modes, so entire repositories and long agent traces fit in one session.
• Day-one open weights: BF16, FP8, INT4, NVFP4, GGUF, and MLX conversions shipped on Hugging Face at launch, plus DFlash draft models for speculative decoding. The repo has reached roughly 119K downloads and 969 likes.
• Runs on one workstation: The model fits and runs on a single NVIDIA DGX Spark with 128GB unified memory.
• Broad serving stack: Poolside's pool CLI, vLLM, SGLang, Ollama, llama.cpp, MLX, and TensorRT-LLM all run it.

Use Cases

• Agentic coding: Built for multi-step tool use, terminal work, and repo-scale edits. On Terminal-Bench 2.1 and SWE-Bench Pro it matches or exceeds models several times its size, including DeepSeek-V4-Flash, NVIDIA Nemotron 3 Ultra, and Thinking Machines Inkling.
• Long-horizon tasks: The 1M window plus interleaved thinking keeps long debugging and refactoring sessions coherent across many tool calls.
• Local deployment: OpenMDW-1.1 weights and 4-bit quantizations make self-hosting on a single DGX Spark practical for privacy-sensitive codebases.

Pricing

Weights are free under OpenMDW-1.1. Hosted access options:

Advantages vs Competitors

1. Efficiency per active parameter: 8B active parameters competing with far larger models on Terminal-Bench 2.1 and SWE-Bench Pro.
2. Truly open for business: OpenMDW-1.1 permits commercial use without a separate agreement, and every major quantization shipped on day one.
3. Local-first at 1M context: Few open models combine a 1M window with single-workstation deployment on a DGX Spark.

Getting Started

1. Try hosted: OpenRouter's free endpoint (256K context) is the fastest way to test it; the Poolside API and Vercel AI Gateway also carry it.
2. Download weights: Grab BF16, FP8, INT4, NVFP4, GGUF, or MLX builds from Hugging Face.
3. Serve locally: Run with the pool CLI, vLLM, SGLang, Ollama, llama.cpp, MLX, or TensorRT-LLM.
4. Control reasoning: Set enable_thinking per request; thinking runs at max when enabled.

FAQ

Is Laguna S 2.1 free for commercial use?

Yes. The OpenMDW-1.1 license covers commercial and non-commercial use, and OpenRouter additionally offers a free hosted endpoint with a 256K context.

What hardware do I need to run it locally?

A single NVIDIA DGX Spark (128GB unified memory) runs the model, and INT4, NVFP4, or GGUF builds reduce the footprint further through llama.cpp, MLX, or TensorRT-LLM.

Does it support images or vision?

No. Laguna S 2.1 is text-to-text only, with a 100,352-token vocabulary and no vision encoder.

Alternatives

• DeepSeek V4 Flash: A larger 284B-A13B open agentic model with MIT weights and cheap API pricing.
• Qwen3.8-27B: A dense 27B Apache 2.0 model for simpler local deployments.

Tips

1. Turn thinking on for hard tasks: enable_thinking runs at max strength and interleaves reasoning between tool calls, which is where the agentic gains come from.
2. Use the DFlash draft models: Speculative decoding with the official DFlash drafters speeds up local serving noticeably.
3. Exploit the full window: With 1M tokens in both modes, keep whole repos and full tool history in context instead of chunking.

Conclusion

Laguna S 2.1 is one of the most practical open coding releases of 2026: 118B MoE with only 8B active, a true 1M-token context, day-one quantizations, and a license that allows commercial use. If you want frontier-adjacent agentic coding on your own hardware or at minimal API cost, it deserves a serious look.