Poolside AI and the Laguna Model Family, Explained

Poolside is a San Francisco AI lab that builds coding models and publishes their weights. Its model line is called Laguna. The company describes its work as building "open-weight foundation models and the systems that refine and improve them."

That single sentence is why business buyers are searching for it. Most strong coding models are locked behind an API. Poolside publishes weights you can download, inspect, and run on your own hardware.

This guide covers what Poolside is, which Laguna variants actually exist, what license each one carries, how to get access, and what it costs. Every figure below comes from Poolside's own pages or its own press release, and every one of them should be re-checked before you sign anything.

What Is Poolside?

Poolside is a private AI lab founded in 2023 and headquartered in San Francisco. It builds foundation models for software engineering and publishes the weights.

The company was co-founded by Eiso Kant and Jason Warner, who serves as co-CEO. Warner was CTO of GitHub before starting the company.

Poolside's own release describes the company as developing open-weight models for software engineering and deploying to government and defense organizations. Warner framed the goal directly in that release: "The West needs open-weight models it can trust, run, and build on."

Poolside also runs satellite offices in Paris and London. The legal headquarters is San Francisco, which matters for buyers who screen vendors on jurisdiction.

• Founded — 2023, headquartered in San Francisco (Poolside press release, July 21, 2026).
• Co-founders — Eiso Kant and Jason Warner; Warner is co-CEO and a former GitHub CTO.
• What it builds — open-weight foundation models aimed at agentic coding and long-horizon software work.
• Where the weights live — the Poolside organization on Hugging Face.
• Who it sells to — developers via API and gateways, plus government and defense organizations.

What the Laguna Model Family Actually Includes

Laguna is Poolside's model line, and it ships in size tiers rather than as one model. As of August 2026, Poolside's own models page lists two current models: Laguna S 2.1 and Laguna XS 2.1.

Poolside's Hugging Face organization also hosts an earlier generation. Laguna M.1 and Laguna XS.2 remain published there alongside the 2.1 models.

The naming trips people up, so read it as two axes. The letter is the size class — M, S, XS. The number after it is the generation, so 2.1 is newer than M.1.

• Laguna S 2.1 — 118B total parameters with 8B active per token, and a 1,048,576-token context window. Released July 21, 2026.
• Laguna XS 2.1 — 33B total parameters with 3B active per token, and a 262,144-token context window.
• Laguna M.1 — 225B total parameters with 23B active per token, and a 262,144-token context window. Earlier generation, still published.
• Laguna XS.2 — the earlier small tier, still listed in Poolside's Hugging Face organization.
• Architecture — these are sparse mixture-of-experts models, so the active parameter count, not the total, drives inference cost.

What License Do the Laguna Weights Carry?

The license is not the same across every Laguna model, and this is the detail most write-ups get wrong. Check the model card for the exact file you plan to download.

As of August 2026, the Hugging Face cards for Laguna S 2.1 and Laguna XS 2.1 both show OpenMDW-1.1. The card for the earlier Laguna M.1 shows Apache 2.0.

Poolside's own models page describes its licensing as a range, covering "permissive licenses from OpenMDW to Apache 2.0." It does not publish a single blanket license for the whole family.

• Laguna S 2.1 — OpenMDW-1.1 on its Hugging Face model card.
• Laguna XS 2.1 — OpenMDW-1.1 on its Hugging Face model card.
• Laguna M.1 — Apache 2.0 on its Hugging Face model card.
• Practical rule — treat "open-weight" as a description of availability, not as a specific set of rights. The license text decides what you may do.
• Before deployment — have whoever signs off on third-party licenses read the actual license file, not a summary of it.

How Poolside Positions Laguna Against Coding Assistants

Poolside sells a model, not an editor. Laguna is trained for agentic coding — running multi-step work through tools rather than autocompleting a line.

That is a different product category from an IDE-first assistant. Cursor gives you an editor with a model inside it. Poolside gives you weights and an API, and you supply the harness.

Poolside says the models were trained with reinforcement learning inside its own agent harness. The models support interleaved thinking between tool calls, which you can switch on or off per request.

• Built for — long-horizon software tasks: reading a repository, planning, editing across files, running tools.
• Not built for — drop-in editor autocomplete without an agent framework around it.
• Tool calling — native, with reasoning preserved across tool calls.
• Thinking mode — can be enabled or disabled per request, so you pay for reasoning only when a task needs it.
• Who assembles the workflow — you do, or your vendor does. The model is the component, not the finished product.

What Benchmark Numbers Has Poolside Published?

Poolside published three headline coding scores for Laguna S 2.1 in its release announcement. These are vendor-reported numbers, which is the normal caveat for any launch.

The claim underneath them is a size argument. Poolside says the 118B model matches or exceeds models several times larger on agentic coding tasks.

Poolside also disclosed the training run: under four weeks end to end on 4,000 H200 GPUs, using an internal platform it calls Model Factory.

• Terminal-Bench 2.1 — 70.2% (Poolside, July 21, 2026).
• SWE-bench Multilingual — 78.5% (Poolside, July 21, 2026).
• SWE-bench Pro — 59.4% (Poolside, July 21, 2026).
• Training — under four weeks on 4,000 H200 GPUs, per Poolside's own release.
• Verify before you decide — run your own repositories through the model. A published benchmark is a starting filter, not an evaluation.

How to Get Access to Laguna

There are three practical paths, and they suit very different teams. Download the weights, call Poolside's API, or route through a third-party gateway.

Downloading is the path that makes Laguna interesting to regulated buyers. The weights sit on Hugging Face, and the model cards list launch-day support for common serving stacks.

If you do not want to run GPUs, the hosted paths get you to a working call in minutes. Poolside runs its own API platform, and the models are also listed on OpenRouter and the Vercel AI Gateway.

• Self-host — pull the weights from the Poolside organization on Hugging Face.
• Serving stacks — the model cards cite vLLM, SGLang, Transformers, TensorRT-LLM, llama.cpp, and Ollama support at launch.
• Poolside API — an OpenAI-compatible chat API through Poolside's own platform.
• Gateways — OpenRouter and the Vercel AI Gateway both list Laguna models.
• Local hardware — Poolside says Laguna S 2.1 runs on a single NVIDIA DGX Spark, and the XS 2.1 card cites machines with 36GB of RAM.

What Does Laguna Cost?

Poolside does not publish a pricing page on its own site as of August 2026. The weights themselves are free to download; what you pay for is the compute you run them on.

For the hosted path, third-party gateways publish rates. OpenRouter listed Laguna S 2.1 at $0.09 per million input tokens and $0.18 per million output tokens on August 2, 2026, with Laguna XS 2.1 at $0.06 and $0.12.

Gateway pricing changes without notice, so treat those figures as a snapshot rather than a rate card. Check the live listing before you build a budget on it.

• Weights — no license fee to download from Hugging Face; your cost is hardware and operations.
• Poolside API — no public rate card on poolside.ai as of August 2026. Contact the company for terms.
• OpenRouter listing (August 2, 2026) — Laguna S 2.1 at $0.09 in / $0.18 out per million tokens.
• OpenRouter listing (August 2, 2026) — Laguna XS 2.1 at $0.06 in / $0.12 out per million tokens.
• Free tiers — OpenRouter also lists free 256K-context endpoints for both 2.1 models, which are fine for evaluation and wrong for production planning.

Who Laguna Actually Fits — and Who It Does Not

Laguna fits teams that need the weights in their own environment. That is usually a compliance requirement, a data-residency requirement, or a cost problem at high volume.

It fits much less well if you want a finished product. There is no Poolside IDE that competes with a mature coding assistant, and assembling an agent harness is real engineering work.

Layer3 Labs advises businesses on model selection and self-hosting tradeoffs, and one pattern repeats in nearly every engagement. Teams shortlist on benchmark scores, then discover the deciding constraint was hardware they already own and a license their legal review will actually approve. A small-active-parameter model that runs on one machine changes that conversation more than a few points on a leaderboard does.

• Good fit — regulated teams that cannot send source code to a hosted API.
• Good fit — high-volume coding automation where per-token API cost has become the bottleneck.
• Good fit — teams already running vLLM or SGLang who want to add a coding model to an existing stack.
• Poor fit — a small team that wants an editor today and has nobody to run inference.
• Poor fit — buyers who need a named compliance certification. Poolside publishes little in the way of formal attestations.

Limits and Open Questions to Check Yourself

The published facts are good; the unpublished ones matter too. Three gaps are worth raising with Poolside directly before a commitment.

First, training data. Poolside does not publish a detailed disclosure of what the models were trained on, which some legal reviews will treat as a blocker for code generation.

Second, support and lifecycle. There is no published statement on how long a given Laguna version stays available or supported. Third, formal certifications — check for anything your procurement process requires rather than assuming it exists.

• Training data — no detailed public disclosure. Ask directly if your legal review requires provenance.
• Version lifecycle — no published support window. Downloaded weights are yours, which is part of the point.
• Certifications — verify against your own procurement checklist; do not assume.
• Benchmarks — vendor-reported. Reproduce on your own code before you decide.
• License — read the file on the specific model card, because it differs across Laguna variants.

What you need to run Laguna S 2.1 yourself

Laguna S 2.1 needs real memory, but it is within reach of a high-end workstation or a couple of professional GPUs — and many teams simply rent instead of buying. Match the path below to whether you want to own the hardware or pay by the hour.

To put Laguna S 2.1 to work once it is live, connect a coding client like Cursor (via OpenRouter) or a local runner such as Ollama.

Apple Mac Studio (M4 Max, 128GB)

A single quiet on-prem box

View on Amazon →

NVIDIA RTX 6000 Ada (48GB)

Power users who want hardware they own

View on Amazon →