Laguna S 2.1
Laguna S 2.1 is Poolside's open-weight coding and reasoning model released on July 21, 2026. It has 118 billion total parameters, activates about 8 billion parameters per token, supports a 1,048,576-token context window, and is published as `poolside/Laguna-S-2.1` under the OpenMDW-1.1 license.
Laguna S 2.1 gives teams a comparatively sparse model for long agent sessions without requiring a closed frontier API. Its value is not just the headline context size: Poolside publishes official BF16 and quantized weights, serving guidance, reasoning controls, known tool-use limitations, and evaluation trajectories that make the model easier to test responsibly.
Poolside's launch post and official Hugging Face model card are the factual authorities for architecture, context, license, weights, serving, hosted access, limitations, and vendor-reported evaluations. Hacker News is used only as evidence that developers want to understand local deployment, coding quality, long-context behavior, and comparison tradeoffs.
- Long-horizon repository work with an agent harness that preserves reasoning between tool calls.
- Self-hosted coding-model evaluation with vLLM, SGLang, TRT-LLM, or llama.cpp.
- Million-token experiments where the runtime and hardware can sustain the requested context.
- Hosted coding-agent tests through a provider that publishes the exact context and price for its endpoint.
Poolside publishes the main BF16 checkpoint plus FP8, NVFP4, INT4, GGUF, MLX, and DFlash variants. The official card documents Transformers, vLLM, SGLang, TRT-LLM, and llama.cpp paths, while the launch post lists hosted access through OpenRouter, Baseten, Vercel AI Gateway, and Poolside chat.
- Official model ID: `poolside/Laguna-S-2.1`.
- Architecture: 118B total parameters, about 8B active, 256 routed experts plus one shared expert.
- Context: 1,048,576 tokens on the official checkpoint; provider endpoints may expose smaller limits.
- License: OpenMDW-1.1, not Apache 2.0.
Poolside reports strong results on terminal and software-engineering evaluations, but those scores were produced through specified harnesses and should be treated as vendor-reported until independently reproduced. The model card also warns about first-call tool-schema mistakes, nested JSON escaping errors, and overthinking, which can matter more than a leaderboard score in a production agent loop.
The BF16 checkpoint is roughly 236 GB of weights and needs multiple GPUs. Official quantizations reduce memory requirements, but a quantized model is not automatically a consumer-laptop model. Teams should validate the exact format, runtime branch, tool parser, reasoning parser, context target, and throughput before choosing a local deployment.
Laguna S 2.1 FAQ
Common questions about Laguna S 2.1.
Is Laguna S 2.1 open weight?+
Yes. Poolside publishes the main checkpoint and several official quantized variants on Hugging Face under OpenMDW-1.1. Check that license directly for your intended use instead of assuming the model uses Apache 2.0.
How much context does Laguna S 2.1 support?+
The official model card specifies a 1,048,576-token context window. Hosted providers may expose different limits: Poolside says OpenRouter's free endpoint uses 256K while a dedicated paid endpoint exposes the full 1M context.
Can Laguna S 2.1 run locally?+
Yes, but hardware requirements depend on the format. Poolside estimates roughly 236 GB for the BF16 weights and publishes lower-precision variants plus GGUF and MLX conversions. Verify the chosen runtime and quantization on your actual hardware before treating it as practical.
Is Laguna S 2.1 better than Kimi K3 or GPT-5.6?+
That is task-dependent. Poolside publishes vendor-run coding and agent benchmarks, but model size, harness, context, tool reliability, latency, cost, and review burden differ. Compare the same repository tasks and acceptance tests before choosing a model.