Poolside Open-Weighted a 118B Coding Model That Runs on One Desktop. It Beats Bigger Open Models and Still Skips the Frontier Fight.

TL;DR

Poolside just open-weighted Laguna S 2.1, a 118-billion-parameter coding model that lights up only 8 billion parameters per token, carries a 1,048,576-token context, and fits on a single desktop once you quantize it. It scores 70.2% on Terminal-Bench 2.1 and 59.4% on SWE-Bench Pro, which beats open models several times its size but sits well behind the closed frontier and China's Kimi K3. The kicker: the lab shipping what it calls the West's most capable open-weight model is the same one whose $2 billion funding round fell apart earlier this year over doubts it could compete at all.

What Poolside actually shipped

Laguna S 2.1 is a mixture-of-experts model: 118B parameters total, but a router picks roughly 8B of them to run on any given token. Think of it as a big consulting firm where only a couple of specialists get pulled into each meeting. You pay to keep everyone on payroll (the weights live in memory) but you only pay the compute of the two people who actually talk.

That gives you the knowledge footprint of a large model at the inference cost of a small one. Concretely: 256 routed experts with a top-10 selection plus one shared expert, 48 layers, mostly sliding-window attention, and a full 1M-token context with separate thinking and no-thinking modes. Poolside says it trained the model over roughly two months (May 22 to July 21) on about 4,000 NVIDIA H200 GPUs using its internal Model Factory, as a scale-up from the smaller Laguna XS it put out a few weeks earlier.

The design goal was not raw IQ. Co-head of applied research Pengming Wang framed it as behavior, not brains:

What we've done in this model is not necessarily add more intelligence, but improve the behaviors that lead to a more capable model: more verification, less taking things for granted.

In other words, they tuned it to check its work instead of confidently guessing, which is exactly the failure mode that makes coding agents infuriating.

The honest headline: it is not the frontier

Poolside's own numbers make the ceiling clear. On Terminal-Bench 2.1, Laguna S 2.1 lands at 70.2%. Kimi K3 and Claude Fable 5 both sit around 88%. That is a gap of nearly 18 points, and Poolside does not hide it.

So why does this matter? Because the comparison Poolside cares about is not against the closed frontier, it is against everything you can actually download and run. There, the story flips. Laguna S 2.1 matches or beats open models with far more active parameters, including DeepSeek's V4 line, NVIDIA's Nemotron 3 Ultra, and Thinking Machines' Inkling. Its 78.5% on SWE-Bench Multilingual even nudges past Qwen 3.7 Max at 78.3%. Punching several weight classes up while activating 8B parameters is the whole pitch.

The number that matters: 8 billion active

Here is where a builder should sit up. A model with 8B active parameters and aggressive quantization runs on a single NVIDIA DGX Spark, the 128GB desktop box, rather than a rack of accelerators. The BF16 checkpoint is about 236GB and still wants multiple GPUs, but Poolside also ships FP8, INT4, NVFP4, and GGUF variants, so the practical footprint drops fast.

Read the shape of that chart, not just the heights. Laguna is best at multilingual repo edits (78.5%) and solid on terminal work (70.2%), but drops to 49.7% on Toolathlon and 40.4% on DeepSWE, both long-horizon, tool-heavy tasks. The model is a sharp software engineer that gets shakier the longer the leash. For local autocomplete, refactors, and bounded agent loops on your own codebase, that trade is more than fine. For a fully autonomous overnight agent, keep a human near the kill switch.

Open-weight, and it means it

The license is the quiet headline for anyone who has been burned by "open" models that ban commercial use. Laguna S 2.1 ships under OpenMDW-1.1, the Linux Foundation's permissive Open Model, Data and Weights license, which grants royalty-free rights to use, modify, and redistribute the weights without the usage carve-outs common in bespoke model licenses.

The runtime story is equally unfussy. Poolside lists support across vLLM, SGLang, TRT-LLM, llama.cpp, and Transformers, plus a trained draft model for speculative decoding. Weights, quants, and inference paths that all exist on day one is a rarer thing than it should be.

The West's open-weight gap, priced in a failed round

Context is what makes this launch interesting rather than just competent. Poolside was founded by ex-GitHub CTO Jason Warner and Eiso Kant, and raised a $500M Series B in October 2024 at roughly a $3 billion valuation, with NVIDIA and eBay among the backers. Earlier in 2026 it was reportedly closing a $2 billion Series C near a $12 billion valuation, with NVIDIA slated to anchor up to $1 billion. That round collapsed. The Financial Times reported investors doubted Poolside could credibly train models to compete with Anthropic, OpenAI, and Google DeepMind at the frontier, and NVIDIA declined to lead a rescue.

Laguna S 2.1 is the reply to that doubt, and it is a revealing one. It does not reach the frontier. Instead it plants a flag on ground the frontier labs have largely abandoned: open weights, small active footprint, runs on your desk, permissively licensed. Warner's framing says the quiet part out loud.

The West needs open-weight models it can trust, run, and build on.

That is a pointed sentence in a year when the most capable downloadable coding models have overwhelmingly come from Chinese labs like DeepSeek, Qwen, and Moonshot's Kimi. Poolside is not arguing it beat them on the leaderboard. It is arguing the West should have a credible entry at all, and that efficiency plus a real license is a defensible place to stand when you have conceded the top of the chart.

Should you actually run it?

If you have a DGX Spark or a couple of consumer GPUs and you want a strong, private, permissively licensed coding model that never phones home, this is now one of the best options you can self-host, full stop. If you are chasing the absolute best agentic coding accuracy and cost is no object, the closed frontier and Kimi K3 are still ahead. Most builders live between those poles, which is exactly the gap Laguna is aimed at.

Key Takeaways

• 118B total, 8B active. A mixture-of-experts design gives Laguna S 2.1 big-model knowledge at small-model inference cost, with a 1M-token context.
• It runs on one desktop. Quantized (FP8/INT4/NVFP4/GGUF), it fits on a single NVIDIA DGX Spark; the BF16 checkpoint is about 236GB.
• Strong for its size, not the frontier. 70.2% Terminal-Bench 2.1 and 59.4% SWE-Bench Pro beat much larger open models but trail Kimi K3 and Claude Fable 5 by roughly 18 points.
• Genuinely open. OpenMDW-1.1 licensing plus day-one vLLM, SGLang, llama.cpp, and Transformers support, all on Hugging Face.
• Best at bounded coding. Sharp on repo edits and terminal tasks, weaker on long-horizon tool use (49.7% Toolathlon, 40.4% DeepSWE).
• The stakes are strategic. A Western lab whose $2B round collapsed on frontier doubts is betting on open weights and efficiency instead of the top of the chart.

Sources: Poolside blog: Introducing Laguna S 2.1, Laguna S 2.1 model card (Hugging Face), The Next Web, Terminal-Bench 2.1, SWE-Bench, OpenMDW license