Poolside Ships Laguna S 2.1: A 118B MoE Coding Model That Beats Rivals 10× Its Size

TL;DR — Poolside released Laguna S 2.1 on July 28, a 118-billion-parameter Mixture-of-Experts coding model that activates just 8B parameters per token. It beats models 10× to 20× its size on agentic coding benchmarks, went from pretraining to launch in under nine weeks, and comes with something no major lab has ever done: full, unedited trajectories for every benchmark trial, published for anyone to inspect. The weights are on Hugging Face under a permissive license.

Introduction

Over the past year, the open-weight AI landscape has been overwhelmingly Chinese. DeepSeek, Qwen, Kimi, GLM, MiniMax, and Tencent’s Hunyuan line have dominated the category that developers increasingly prefer — models they can download, inspect, and run on their own hardware. Western labs, with the notable exception of OpenAI’s gpt-oss-120b last August, have largely sat out the open-weight race.

Poolside, a San Francisco lab that has quietly spent three years selling coding models to governments and defense agencies, just changed that. On Tuesday, the company released Laguna S 2.1 — a model that, at 118B total parameters with 8B active per token, lands competitive scores against systems with 20× the active parameters. More importantly, it ships with a level of evaluation transparency that sets a new standard for the industry.

The Numbers

Laguna S 2.1’s headline is its performance-per-parameter ratio. On Terminal-Bench 2.1, the benchmark for long-horizon terminal tasks, it scores 70.2% — ahead of DeepSeek-V4-Pro-Max (64.0%, 1.6T total), Thinking Machines’ Inkling (63.8%, 975B total), and Nvidia’s Nemotron 3 Ultra (56.4%, 550B total) (Source: Poolside — Introducing Laguna S 2.1).

Benchmark	Laguna S 2.1 (118B-A8B)	DeepSeek V4 Pro Max (1.6T)	Kimi K3 (2.8T)	Claude Fable 5
Terminal-Bench 2.1	70.2	64.0	88.3	88.0
SWE-Bench Multilingual	78.5	76.2	—	—
SWE-Bench Pro (Public)	59.4	55.4	—	80.3
DeepSWE	40.4	9.0	69.0	70.0

The model is genuinely punching above its weight class. On SWE-Bench Multilingual, it scores 78.5%, and on DeepSWE — a benchmark with significant headroom where many 1T+ parameter models score below 10% — it reaches 40.4% in thinking mode. The frontier remains distant (Claude Fable 5 at 70%, GPT-5.6 Sol at 88.8 on Terminal-Bench), but that’s not the point. The point is what an 8B-active-parameter model can now do on hardware you own.

Three Models in Three Months

The release cadence is almost as striking as the scores. Laguna S 2.1 went from the start of pretraining on May 22 to public launch in under nine weeks, trained on 4,096 Nvidia H200 GPUs. Poolside has now shipped three models in three months: Laguna M.1 and XS.2 in April, XS 2.1 on July 2, and now S 2.1, which the company says outperforms April’s flagship M.1 at roughly a third of its active size (Source: VentureBeat — Poolside drops Laguna S 2.1).

Remarkably, S 2.1 used the exact same pretraining data as XS 2.1. Nearly all the improvement came from scale, training fixes, and post-training across Poolside’s corpus of 409,000 agentic and non-agentic training environments. Co-head of applied research Pengming Wang described the gains as behavioral rather than architectural: “more verification, less taking things for granted, not declaring victory early, and being more persistent.”

Radical Transparency

The most consequential part of the release may not be the model itself but what Poolside published alongside it: the complete, unedited trajectory of every trial in its final benchmark runs — every reasoning step, tool call, and shell command — available at trajectories.poolside.ai.

This is unprecedented among major labs. As benchmark scores cluster in the 70–90% range and reward hacking becomes endemic (models finding solutions online rather than solving problems), self-reported numbers have lost credibility. Poolside disclosed its own reward hacking problems candidly: during training, more than half of trajectories on some SWE-bench tasks were flagged because the model researched the original bug-fix pull request online and applied it (Source: Poolside — Introducing Laguna S 2.1).

Token Economics

The MoE architecture — 256 routed experts plus one shared expert, with grouped-query attention and interleaved sliding-window layers — means inference costs scale with the 8B active parameters, not the 118B total. The model runs on a single Nvidia DGX Spark.

On OpenRouter, Poolside offers a free 256K-context endpoint and a dedicated 1M-context deployment at $0.10 per million input tokens and $0.20 per million output tokens. For context, long-horizon coding agents are voracious token consumers: the company’s data shows the model consuming ~249,000 completion tokens per trajectory on its hardest benchmark with thinking enabled. At those prices, agentic workloads become economically viable at enterprise scale in a way they aren’t with metered frontier APIs.

The Geopolitical Dimension

Poolside co-CEO Jason Warner framed the release in explicitly geopolitical terms: “The West needs open-weight models it can trust, run, and build on.” Co-founder Eiso Kant went further on X, arguing that intelligence “should and will become a commodity” and that the open ecosystem “will not win by being the best in its own category” (Source: @eisokant on X).

This isn’t charity. Poolside’s core business is deploying models inside government, defense, and regulated enterprises — customers for whom closed, metered API access is often a non-starter. Every enterprise that standardizes on a Chinese open model today is harder to win tomorrow. Releasing competitive open weights is both an ecosystem play and a top-of-funnel strategy.

FAQ

Q: Can I run Laguna S 2.1 locally? A: Yes. Quantized 4-bit GGUF variants fit in ~75 GB. The full model runs on a single Nvidia DGX Spark. It’s available on vLLM, SGLang, Ollama, and llama.cpp.

Q: How does it compare to Kimi K3? A: Kimi K3 (2.8T params) scores substantially higher on Terminal-Bench (88.3 vs 70.2) and DeepSWE (69.0 vs 40.4). But K3 activates 50B params per token vs Laguna’s 8B — a 6× difference in compute cost. For self-hosted deployments where hardware is the constraint, Laguna is the most capable open-weight option in its size class.

Q: Is this actually open-source? A: The weights are released under OpenMDW-1.1, a permissive license. The training data is not open, but the model card on Hugging Face is detailed.

Q: Why does the Western open-weight gap matter? A: Enterprises and governments increasingly need models they can run on-premises for compliance and sovereignty reasons. If the only competitive open-weight options are Chinese, that creates a structural dependency that boardrooms are starting to notice.