Poolside · 2026-07-21 · major
Poolside Laguna S 2.1 — 118B open-weight coding MoE with 8B active
Poolside ships Laguna S 2.1, a 118B/8B-active open-weight coding MoE with a 1M-token context. Weights land on Hugging Face under the OpenMDW-1.1 license; API access is $0.10/$0.20 per 1M tokens on OpenRouter.
Poolside ships Laguna S 2.1 — a 118B/8B-active open-weight coding MoE with a 1M-token context, pitched as the West's answer to DeepSeek and Qwen.
Key specs
| Active params | 8B |
|---|---|
| Context window | 1M tokens |
| Total parameters | 118B |
Quick facts
| Maker | Poolside |
|---|---|
| Size | 118B total parameters (Mixture-of-Experts), 8B active per token |
| Context window | 1,048,576 tokens (1M) |
| License | OpenMDW-1.1 (open weights, free commercial + non-commercial use) |
| Training | ~9 weeks from start to launch on ~4,000 NVIDIA H200 GPUs |
| Availability | Weights on Hugging Face; API on OpenRouter; hosted chat at chat.poolside.ai; runs on a single NVIDIA DGX Spark |
| Price (input / output) | $0.10 / $0.20 per 1M tokens on OpenRouter |
Benchmarks
Pricing
| Input · OpenRouter list price | $0.10 / 1M tokens |
|---|---|
| Output · OpenRouter list price | $0.20 / 1M tokens |
| Self-hosted · Open weights under OpenMDW-1.1 on Hugging Face | Free |
What is it?
Laguna S 2.1 is Poolside's new open-weight coding foundation model, released July 21, 2026 with weights on Hugging Face under the Linux Foundation's OpenMDW-1.1 license. The model targets long-horizon software engineering work — building systems, refactoring, running through hundreds of thousands of tokens of a single session — and comes with BF16, FP8, NVFP4, INT4, and GGUF checkpoints plus a DFlash draft model for speculative decoding.
How does it work?
Under the hood, Laguna S 2.1 is a 118B-parameter Mixture-of-Experts with 8B active parameters per token, 256 routed experts (top-10 router) plus one shared expert, 48 layers arranged in a 1:3 ratio of global to sliding-window attention, and a 1,048,576-token context. Poolside says it was trained end-to-end in under 9 weeks on ~4,000 NVIDIA H200 GPUs via its in-house Model Factory pipeline, and the sparse activation lets the full model fit on a single NVIDIA DGX Spark box.
Why does it matter?
Chinese labs like DeepSeek and Alibaba have owned the open-weight coding tier for over a year; Laguna S 2.1 is the first Western entry that Poolside argues can trade blows with them. On Terminal-Bench 2.1 (70.2%) and SWE-Bench Pro (59.4%), the model beats or ties DeepSeek-V4-Flash, NVIDIA's Nemotron 3 Ultra, and Thinking Machines' Inkling despite running many times fewer active parameters, and $0.10 / $0.20 per million tokens on OpenRouter makes it roughly an order of magnitude cheaper than closed frontier coding APIs.
Who is it for?
Engineers who want a self-hostable, permissively-licensed coding model — plus AI infra teams and enterprises that cannot ship code out to a closed API.
Frequently asked questions
- Is Laguna S 2.1 open source?
- Yes — Poolside publishes Laguna S 2.1 weights on Hugging Face under the OpenMDW-1.1 license, which permits free commercial and non-commercial use and modification. Poolside also ships BF16, FP8, NVFP4, INT4, GGUF, and a DFlash draft checkpoint for speculative decoding, so the same release covers full-precision inference and quantized local runs.
- What hardware do I need to run Laguna S 2.1 locally?
- Poolside claims Laguna S 2.1 is small enough to run on a single NVIDIA DGX Spark thanks to its 8B active parameters per token. Community-quantized GGUF and INT4 checkpoints on Hugging Face bring the memory footprint down further for smaller multi-GPU rigs; the full 1M-token context still benefits from the DFlash draft model for faster decoding.
- How does Laguna S 2.1 compare to DeepSeek and Qwen coding models?
- Poolside positions Laguna S 2.1 as the West's most capable open-weight coding model, filling a gap that DeepSeek and Alibaba's Qwen have held for over a year. On Terminal-Bench 2.1 (70.2%) and SWE-Bench Pro (59.4%), Poolside reports it matches or beats larger models like DeepSeek-V4-Flash, NVIDIA's Nemotron 3 Ultra, and Thinking Machines' Inkling.
- How much does the Laguna S 2.1 API cost?
- Laguna S 2.1 lists at $0.10 per million input tokens and $0.20 per million output tokens on OpenRouter — an order of magnitude cheaper than frontier closed coding models. OpenRouter notes real prices can drop 60–80% further with prompt caching, and the OpenMDW-1.1 license means teams can self-host the weights instead of paying per token.
- What is Poolside's Model Factory that built Laguna S 2.1?
- Model Factory is Poolside's internal pretraining and post-training pipeline that took Laguna S 2.1 from a fresh run to a shipped model in under 9 weeks on ~4,000 NVIDIA H200 GPUs. Poolside frames it as the reason the West can now ship agentic coding models at Chinese-lab cadence rather than only reacting to DeepSeek and Qwen drops.
Try it
huggingface.co/poolside/Laguna-S-2.1 · OpenRouter model id: poolside/laguna-s-2.1 · hosted chat: chat.poolside.ai