Laguna S 2.1 Released: Poolside Drops Open-Weight Coding Model That Beats Rivals 14x Its Size

On July 21, 2026, an open-weight coding model that activates only 8 billion parameters per token posted a higher Terminal-Bench 2.1 score than DeepSeek V4 Pro Max — a 1.6-trillion-parameter system with 49 billion active parameters per token.

That model is Laguna S 2.1, and it was not built by OpenAI, Anthropic, or DeepSeek.

Poolside, a San Francisco AI lab that spent most of its three-year existence quietly selling coding models to governments and defense agencies, released the weights to the public on day one under the permissive OpenMDW-1.1 license. The model went from the start of pre-training to launch in under nine weeks on 4,096 NVIDIA H200 GPUs. Poolside has now shipped three models in three months.

This is not a video generation model or an image generator. Laguna S 2.1 is a text-based agentic coding model — a direct competitor to Claude Code, GPT-5.6, and DeepSeek V4 in the AI-assisted software engineering space. But the architectural efficiency on display has implications far beyond coding. If an 8B-active model can outperform a 1.6T-parameter system by using smarter post-training, the same approach will reshape how we build video generation models like Wan 2.7 and real-time AI systems in the coming years.

By the end of this article, you will know exactly what Laguna S 2.1 is, how its benchmarks stack up against frontier rivals, what it costs to run, and how to access it — free — today.

What Is Laguna S 2.1?

Laguna S 2.1 is an open-weight Mixture-of-Experts model with 118 billion total parameters and 8 billion active parameters per token, designed for agentic coding and long-horizon software engineering work. It supports a context window of up to 1 million tokens in both thinking and non-thinking modes.

What 118B MoE / 8B Active Actually Means

If you have followed model releases in 2026 — from NVIDIA Nemotron 3 Ultra (550B with 55B active) to Kimi K3 (2.8T with 50B active) — you know that total parameter count is not the number that matters. What matters is active parameters per forward pass.

In a Mixture-of-Experts architecture, the model routes each input token through a small subset of specialized expert sub-networks. Only 8 billion parameters fire per token. The other 110 billion sit idle until they are needed.

What this means in practice:

• Inference cost is closer to an 8B model than a 118B model. You pay for 8B active parameters, not 118B.
• Knowledge capacity exceeds any 8B dense model. With 14x more total parameters, the model stores far more specialized capability across experts.
• Self-hosting is feasible on hardware you can buy. A quantized GGUF version (~75GB at 4-bit) fits on a single high-end consumer GPU or a multi-GPU workstation.

Rule of Thumb: If you currently pay for hosted coding models by the token, Laguna S 2.1's MoE design means switching could cut your cost to one-twentieth or less compared to equivalently capable dense models.

Benchmark Performance

The raw numbers tell the efficiency story. Laguna S 2.1 is, by a wide margin, the most capable model in its weight class on every major coding benchmark.

Scores as of July 21, 2026. pass@1 averaged over 3–4 attempts per task.

The headline is the Terminal-Bench 2.1 score of 70.2% in thinking mode. On this benchmark — which tests long-horizon terminal tasks where an agent must persist, debug, and iterate — Laguna S 2.1 sits at #11 on the compiled leaderboard. It beats DeepSeek V4 Pro Max (1.6T total, 49B active) by 6.2 points, Thinking Machines' Inkling (975B, 41B active) by 6.4 points, and NVIDIA Nemotron 3 Ultra (550B, 55B active) by 13.8 points.

But the real signal is on DeepSWE v1.1, where Laguna S 2.1 scores 40.4% while DeepSeek V4 Pro Max scores just 9.0%. DeepSWE tasks are longer-horizon and harder to partially solve — a model cannot luck its way through them. The 31.4-point gap is evidence that Poolside's investment in persistence and verification behaviors, not just raw intelligence, pays off on difficult real-world problems.

The Thinking Mode Differential

Laguna S 2.1 ships with two modes, and the gap between them is substantial:

Thinking mode roughly doubles the model's DeepSWE score and adds roughly 10 points on every benchmark. The cost is higher token usage: thinking trajectories average 129,000 completion tokens on Terminal-Bench compared to 80,000 in non-thinking mode — about 1.6x the output tokens.

Rule of Thumb: For complex multi-file tasks, always use thinking mode. For fast autocomplete or single-function generation, non-thinking mode is the right choice and saves nearly 40% on output tokens. If you are not sure which one you need, start with non-thinking mode — it still scores 60.4% on Terminal-Bench and will catch most simple failures before you need the heavier mode.

A model that outperforms systems with 6x to 20x more active parameters sounds impressive on a leaderboard. What matters for most developers is the next question: what does it actually cost to use?

Pricing and Access — Free and Paid Options

Poolside made access straightforward. Every option below is available today.

Free Access

Paid Access

For comparison: Claude Sonnet 5 costs $2/$10 per million tokens. GPT-5.6 Terra costs $1.50/$15. Laguna S 2.1 on OpenRouter is 20x cheaper on input and 50x cheaper on output than Sonnet 5 — for comparable SWE-Bench Pro performance within a few percentage points.

Self-Hosted: Run It on Your Own Hardware

The weights are on Hugging Face in BF16, FP8, INT4, and NVFP4 formats. Official GGUF conversions are available from Unsloth and Poolside. Supported serving frameworks include vLLM, SGLang, Ollama, and llama.cpp.

The full model at BF16 weighs roughly 235GB. Quantized GGUF variants at 4-bit fit within 75GB — runnable on a dual 4090 or a single A6000 workstation. For machines with 24GB VRAM, Poolside also offers Laguna XS 2.1 (33B total, 3B active), which scores 33.4% on Terminal-Bench and 63.1% on SWE-Bench Multilingual — comparable to the April release Laguna M.1 at roughly a seventh of the active size.

Ecosystem Support: Where You Can Use Laguna S 2.1 Today

Poolside worked with the ecosystem before launch to ensure broad day-one availability. The model is live inside the tools developers already use:

The pool CLI is Poolside's own terminal-based coding agent. It connects directly to the model's 1M context window and thinking capabilities. In pool, toggle thinking per session with /thought-level.

How Laguna S 2.1 Was Trained — The Model Factory

Laguna S 2.1 was not simply a scale-up. Poolside's internal platform, the Model Factory, automated the full pipeline: pre-training, post-training, evaluation, and release with minimal human bookkeeping.

From XS 2.1 to S 2.1 — Same Data, Different Scale

The most revealing detail in Poolside's release: S 2.1 used exactly the same pre-training data as XS 2.1, the 33B model released July 2. Nearly all the improvement came from scale, training code fixes, and especially post-training.

Post-Training in Two Stages

Stage 1 — SFT with synthetic data: Supervised fine-tuning bootstrapped capabilities partly from synthetic data.

Stage 2 — RL reserved for hard tasks: Reinforcement learning was applied only to tasks the model could not yet solve at high pass rates. Rollouts happened across multiple agent harnesses simultaneously, so the model learned behaviors that generalize rather than overfitting to any single scaffold.

The RLCEF Approach

Instead of human preference labels, Poolside used Reinforcement Learning with Code Execution Feedback (RLCEF). The model learns from actual code execution results — whether a test suite passes, whether a build succeeds, whether memory usage drops. The training corpus spans 409,000 agentic and non-agentic environments; within that, 83,000 setups target terminal use cases and 168,000 target standard software engineering workflows.

What Changed in the Training Loop

Three specific improvements distinguish S 2.1 from earlier Laguna models:

1. More generous rollout budgets: Longer timeouts, more tokens per turn, and more turns per task than any prior Poolside model. This is one likely reason for the persistence behavior — the model learned that giving up early was undesirable.
2. Better sandbox infrastructure: RL moved to a new sandboxing service with background process support, selective network blocking to prevent reward hacking, and artifact caching to avoid overloading external services.
3. Multi-harness rollouts: The same prompts are executed in several different agent scaffolds, forcing the model to develop behaviors that carry across environments.

Three Models in Three Months

In three months, Poolside went from their first serious agentic coding model to one that competes with trillion-parameter systems — at roughly one-third the active size of their April flagship.

The training story matters because it isolates where the gains came from. Same pre-training data as XS 2.1. Better post-training. No magic dataset — just RL on hard tasks with real execution feedback. That means the recipe is replicable across model sizes and domains, not a one-off result.

Why This Matters Beyond Coding

Laguna S 2.1 is a coding model, but the architecture story is bigger. The 118B MoE design demonstrates that a model with 8B active parameters — trainable in under nine weeks — can match or exceed dense competitors using 6x or more active parameters. That ratio matters for any domain where inference cost is the bottleneck.

In video generation, Wan 2.7 already uses diffusion transformer architectures that are fundamentally different from text-only MoE models. But the trajectory is the same: efficient architectures that activate only what is needed per token or per timestep will dominate the next generation of generative models. If an 8B-active MoE can outperform a 49B-active dense competitor in coding, a diffusion MoE with partial activation could do the same for video generation — generating 720p clips at the cost of a 480p model.

The immediate practical takeaway: open-weight development velocity in the West is accelerating. Poolside's three-models-in-three-months cadence suggests the gap between open-weight and proprietary models — already narrowing — may close faster than most enterprise buyers expect.

Limitations — What Poolside Disclosed Honestly

The numbers above are compelling, but no model release is complete without an honest accounting of what does not work yet. Poolside published known limitations that most labs bury — and pragmatically, these are the issues you will hit first in production. These matter for anyone evaluating the model for production:

Harness overfitting: The model can stumble on slightly different tool schema definitions in third-party agent harnesses when they resemble but do not exactly match its native harness. This typically resolves with one retry when the harness rejects the invalid tool call.

Nested tool call formatting: When a tool argument expects a JSON array, the model may generate incorrectly escaped or invalid JSON. This is an active area of improvement for the next iteration.

Overthinking on competition math: The model can think for long sequences before making progress on competition-style mathematics problems. There is currently no user-configurable thinking effort control — just on or off.

No configurable thinking effort dial: Unlike some competitors that offer low-medium-high thinking budgets, Laguna S 2.1 ships with two modes only: off and max. Poolside says effort control is coming in a future release.

Production risk if you self-host: The model is permissively licensed for commercial use, but state-of-the-art coding agents can generate destructive commands (file deletions, system modifications) at speed. If you run Laguna S 2.1 in an agent harness with shell access, sandbox it. Poolside themselves use an RL sandbox with selective network blocking and background process isolation. Do the same before pointing this at a production filesystem.

FAQ

Is Laguna S 2.1 a video generation or image generation model?

No. Laguna S 2.1 is a text-only agentic coding model. It generates code, shell commands, and reasoning traces — not images or video. If you are looking for video generation, see the Wan 2.7 complete guide for what is currently the best open-source AI video generation model.

Is Laguna S 2.1 really free?

Yes. The 256K-context endpoint on OpenRouter is free. The 1M-context endpoint costs $0.10/$0.20 per million tokens — roughly 20x cheaper than Claude Sonnet 5. You can also download the full weights and run locally at zero ongoing cost.

Can I run Laguna S 2.1 on my own machine?

The full model at BF16 requires about 235GB of VRAM — a multi-GPU workstation or server. Quantized GGUF files at 4-bit fit within 75GB, runnable on a dual-4090 setup or an A6000. For consumer-grade hardware with 24GB VRAM, use Laguna XS 2.1 (33B total, 3B active) instead.

How does Laguna S 2.1 compare to Wan 2.7?

They are entirely different categories of models. Laguna S 2.1 is a text-only coding LLM — it writes and debugs software. Wan 2.7 is an AI video generation model that creates video clips from text, images, or reference footage. Comparing them directly would be like comparing a code editor to a camera. If you need to generate videos, use Wan 2.7 online platforms. If you need an AI coding assistant, Laguna S 2.1 is one of the best free options available today.

What license does Laguna S 2.1 use?

OpenMDW-1.1, a permissive license that allows commercial use, modification, and redistribution. You can build products on top of it without licensing fees.

Should I use thinking mode or non-thinking mode?

For simple tasks (autocomplete, single-function generation, quick edits), non-thinking mode is faster and still scores 60.4% on Terminal-Bench. For complex multi-file tasks, codebase-level refactoring, or debugging sessions that span dozens of turns, always use thinking mode — the jump from 60.4% to 70.2% on Terminal-Bench and from 16.5% to 40.4% on DeepSWE is too large to ignore.

How does it compare to GPT-5.6 or Claude Fable 5?

The frontier closed models still lead: GPT-5.6 Sol scores 88.8% and Claude Fable 5 scores 88.0% on Terminal-Bench 2.1 versus Laguna S 2.1's 70.2%. But the cost difference is staggering — Laguna S 2.1 on OpenRouter is free for 256K context or up to 50x cheaper at full 1M context. If you are doing high-volume coding work where a few percentage points of benchmark delta translates to negligible real-world difference, the economics favor the open model.

Who is Poolside, and why should I trust this model?

Poolside is a San Francisco-based AI lab founded in 2023. Its primary customers are governments, defense agencies, and regulated enterprises that require on-premise deployment for compliance reasons. The company publishes full evaluation trajectories — every reasoning step, tool call, and shell command behind every reported benchmark score — at trajectories.poolside.ai. This level of transparency is rare among major labs.

---

The Bottom Line

Laguna S 2.1 is not the highest-scoring coding model on the planet — GPT-5.6 Sol and Claude Fable 5 each hold roughly 18 points over it on Terminal-Bench. But it resets the cost-to-capability equation in a way no other recent release has done.

• 8B active parameters deliver performance that demands 49B+ active parameters from competing models — meaning you pay for a small model and get large-model reasoning.
• Free on OpenRouter at 256K context, with a 1M-context endpoint at $0.10/$0.20 per million tokens — 20x to 50x cheaper than frontier closed models.
• Self-hostable on a dual-4090 or single A6000 via quantized GGUF (~75GB at 4-bit). No API key required.
• Poolside's trajectory toward smaller, smarter models — from 23B active in April to 8B active in July with 70% on Terminal-Bench — suggests the next generation of efficient architectures (MoE diffusion, partial-activation video models) is closer than most people think. The same efficiency story will play out in video generation, where Wan 2.7 already leads the open-weight category.

What to do next: Go to OpenRouter, select the free endpoint, and give it a real coding task — not a toy prompt, but something that spans multiple files or requires iterative debugging. The only way to know if 70.2% on Terminal-Bench translates to your workflow is to put the model through a task you actually need done.

---

Laguna S 2.1 is an open-weight coding model, not a video generator. If you are here for AI video generation, Wan 2.7 remains the best open-source option for text-to-video, image-to-video, and reference-to-video generation — try it with a free trial on wan27.org.