On DeepSWE v1.1 - the benchmark designed specifically to reward agents that persist, debug, and iterate rather than guess - Laguna S 2.1 reaches 40.4%, while DeepSeek V4 Pro Max scores just 9.0% on the same test at roughly one-sixth the active parameters. That 31-point gap on a hard long-horizon benchmark is worth sitting with before reading any of the headline numbers.
Poolside, a San Francisco lab that spent most of its existence quietly selling coding models to governments, released the weights publicly on July 21, 2026. The weights shipped under the permissive OpenMDW-1.1 license. The post below covers what is genuinely new, what the benchmark table hides, and what you would actually need to run this model yourself.
What Laguna S 2.1 actually is
Laguna S 2.1 is a 118B total-parameter Mixture-of-Experts model with just 8B activated parameters per token, a context window up to 1M tokens, and support for thinking and no-thinking modes.
The model activates roughly 6.8% of its parameters on any given token. All 118B parameters remain resident in memory, but only ~8B route through the network per step - that sparsity is why a mid-size model can behave like a larger one while staying cheap to serve.
What differentiates Poolside's entry is the training-time claim specifically: under nine weeks on 30 trillion tokens is a notably lean run for a 118B-parameter model, and if that figure holds up to scrutiny, it says as much about Poolside's training infrastructure and data pipeline efficiency as it does about the resulting model's benchmark scores.
Pre-training began on 22 May 2026 on 4,096 NVIDIA H200 GPUs. It is the first Poolside model where reinforcement learning ran in FP8 precision.
S 2.1 used the exact same pre-training data as XS 2.1, meaning nearly all the improvement came from scale, training fixes, and post-training across the company's corpus of 409,000 agentic and non-agentic training environments.
The benchmark table - and what it hides
On the benchmark table Poolside itself published, Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, 78.5% on SWE-bench Multilingual, 59.4% on SWE-Bench Pro, 40.4% on DeepSWE, and 49.7% on Toolathlon Verified.
The headline is competitive. The caveat matters: Laguna's headline scores are Poolside-reported (from its launch materials and Hugging Face card) unless a named third party is cited - treat them as vendor-run until independent labs confirm.
The more interesting number is not Terminal-Bench. It is DeepSWE. DeepSWE tasks are longer-horizon and harder to partially solve - a model cannot luck its way through them. The 31.4-point gap over DeepSeek V4 Pro Max is evidence that Poolside's investment in persistence and verification behaviors, not just raw intelligence, pays off.
A comparison of where Laguna S 2.1 stands against the open-weight field on the benchmarks that distinguish it:
| Model | Terminal-Bench 2.1 | DeepSWE v1.1 | Active params | License |
|---|---|---|---|---|
| Laguna S 2.1 | 70.2% | 40.4% | ~8B | OpenMDW-1.1 |
| DeepSeek V4 Pro Max | 64.0%* | 9.0% | 49B | Mixed |
| NVIDIA Nemotron 3 Ultra | 56.4%* | - | 55B | Apache 2.0 |
| Kimi K2.5 | 50.8%† | - | 32B | Modified MIT |
*Poolside compiled leaderboard figures. †Terminal-Bench 2.0 score from Kimi's paper under non-thinking mode.
Closed frontier models like Claude Fable 5, Claude Opus 5, and Kimi K3 still lead on several absolute benchmarks
- Laguna S 2.1 is not a closed-model replacement. What it is: the strongest open-weight coding-per-dollar option currently confirmed on long-horizon agentic tasks, at a weight class where no Western lab had shipped anything in nearly a year.
What you need to actually run it
At 4-bit (NVFP4 or INT4) the weights need about 59 GB, which fits on a single DGX Spark's 128 GB of unified memory. At FP8 they need about 118 GB, still within a single Spark or a single H200. At BF16 they need about 236 GB, which calls for two linked Sparks or a multi-GPU datacenter node.
Poolside publishes weights in BF16, FP8, INT4, and NVFP4, along with official GGUF and MLX conversions and DFlash draft models.
OpenCode added support the same day as the launch, making Laguna S 2.1 free to use with the full 1M context window through one of 2026's default open-source agent harnesses - no separate API key or hosting setup required to start using it.
The meaningful split is between three deployment options:
- OpenRouter API at $0.10/$0.20 per million tokens - fastest start, no hardware. Good for evaluation.
- Single DGX Spark (INT4, ~59 GB) - the practical self-hosted path for a small team. Stays within one machine. One caveat: INT4 introduces quantization loss; Poolside has not published a head-to-head INT4 vs BF16 accuracy delta.
- Multi-GPU node (BF16, ~236 GB) - full fidelity, expensive infrastructure. Realistic only if your threat model requires that no tokens leave your control.
A teammate like Beagle running against an in-house Laguna S 2.1 deployment would keep all context - code, PR history, architecture notes - on your own network. The model handles the reasoning; your infrastructure handles the compliance requirement.
Beagle in action#engineering, 10:42amWhat is genuinely new versus incremental
Poolside's press release frames Laguna S 2.1 explicitly as a response, noting that the model occupies a size class into which no Western lab had released open weights in 11 months - since OpenAI's gpt-oss-120b last August.
The genuinely new things, stated plainly:
- DeepSWE performance at this active parameter count. No previously public model reaches 40% on DeepSWE v1.1 with only 8B active parameters. This is not benchmark cherry-picking: DeepSWE is a hard multi-step eval where partial credit is minimal.
- US-jurisdiction permissive license. US buyers spent the last two years reaching for Chinese open-weight models on cost: DeepSeek, Qwen, and Kimi were simply the strongest weights you could download for the money. Laguna S 2.1 is the first model in this weight class to offer a comparable agentic coding score under a US-lab license.
- Thinking/no-thinking toggle with 1M context. The combination - not thinking mode alone, not 1M context alone - is what makes it usable as a drop-in for long-horizon agent tasks without having to manage context truncation separately.
The incremental or overstated things:
- The math story. The model independently discovered a proof to Erdős problem #397, but this is an independent re-discovery: a proof was found earlier in January 2026 by GPT-5.2 Pro, and Poolside is confident the result was not influenced by the previous result given the model's November 2025 knowledge cutoff. It is an interesting internal test. It is not a claim about frontier math capability.
- The "beats rivals 10x its size" framing. It beats some rivals on some benchmarks. On SWE-Bench Pro (59.4%), it trails Kimi K2.6 (58.6% is close, actually competitive) and is behind GLM-5.2 (81.0% on Terminal-Bench 2.1) by a large margin. The right framing is parameter efficiency on long-horizon tasks, not a global performance crown.
Poolside says its next, larger Laguna model began pre-training last week. The release cadence - the April dual release of Laguna M.1 and XS.2, the July 2 refresh of XS 2.1, and now S 2.1
- is unusually fast for a lab of Poolside's size. Whether the Model Factory infrastructure can sustain this as models scale is the open question.
Laguna S 2.1 open-weight coding model: common questions
What is Poolside Laguna S 2.1?
Laguna S 2.1 is an open-weight, Mixture-of-Experts coding model with 118B total parameters and roughly 8B active parameters per token. It ships under the OpenMDW-1.1 license, supports a 1M-token context window, and scores 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE v1.1 - both vendor-reported figures as of July 21, 2026.
How does Laguna S 2.1 compare to DeepSeek V4 Pro Max?
On Terminal-Bench 2.1, Laguna S 2.1 scores 70.2% vs DeepSeek V4 Pro Max's 64.0% at roughly one-sixth the active parameters. The bigger gap is on DeepSWE v1.1: 40.4% vs 9.0%. DeepSeek V4 Pro Max still leads on some general reasoning benchmarks; this is not a universal win for Laguna.
Can I run Laguna S 2.1 on a single GPU?
At INT4 quantization (NVFP4 or INT4), the checkpoint requires about 59 GB of memory, which fits on a single NVIDIA DGX Spark with 128 GB of unified memory. FP8 needs ~118 GB (one H200). BF16 needs ~236 GB, requiring two linked Sparks or a multi-GPU node.
Is Laguna S 2.1 benchmark data independently verified?
As of publication, all headline scores are vendor-reported from Poolside's launch materials and Hugging Face model card. Poolside states it takes the maximum of its own scores, benchmark-author leaderboards, or Artificial Analysis third-party scores where available. Independent external reproduction is still limited - verify before committing budget.
What license does Laguna S 2.1 use?
The model ships under OpenMDW-1.1, a permissive license that allows commercial deployment. It is different from Apache 2.0 - read the model card for the exact variant you intend to deploy, since license terms vary across the Laguna family.