← Back to feed

Poolside released Laguna XS 2.1 on July 2, offering the model as a free download on Hugging Face and a free tier on OpenRouter — the lowest-friction access point the lab has offered. Developers running the predecessor, Laguna XS.2, have until July 9 before it disappears from OpenRouter and the Poolside API. XS.2 will continue as a dedicated deployment option on Baseten for teams running it on private infrastructure.

What Changed

The model ships with DFlash, a new speculative decoding mechanism that roughly doubles inference speed in local deployments. DFlash is the most specific technical improvement Poolside is citing for the 2.1 release. The model handles multi-step software engineering tasks on a single GPU, positioning it for the local agentic coding workflow that developers building outside the major cloud IDEs increasingly prefer.

Licensing also changed. Laguna XS 2.1 ships under OpenMDW-1.1, a new permissive licence designed specifically for AI model weights. Poolside describes it as giving enterprise legal teams cleaner standing than Apache 2.0, which was written before the current class of large foundation models existed and has accumulated ambiguity around the specific liability and derivative work questions that enterprise AI deployments generate.

For developers already on the paid XS.2 tier, the upgrade is seamless: pricing is matched and only the model ID changes from poolside/laguna-xs.2 to poolside/laguna-xs-2.1. No other integration changes are required.

The Context: A Failed $2B Series C

The free release strategy makes more sense in the context of what happened in April.

Poolside was founded in 2023 by Jason Warner, former CTO of GitHub, and co-founder Eiso Kant. The company raised a $500 million Series B in October 2024 at a $3 billion valuation, led by Bain Capital Ventures, with a separate commitment of up to $1 billion from NVIDIA as part of a larger round expected to close shortly after.

That round fell apart. In April 2026, DataCenterDynamics reported that Poolside’s anchor-tenant agreement with CoreWeave for a planned 2-gigawatt Texas data campus had collapsed, and the associated $2 billion Series C — including the NVIDIA commitment — failed to close after investors raised concerns about Poolside’s ability to train models competitive at the frontier against Anthropic, OpenAI, and Google DeepMind.

The free Laguna XS 2.1 release is the first public signal of how Poolside is repositioning. Rather than competing on frontier capability — a race that requires sustained compute access at the scale of Stargate or Colossus — the lab is building adoption through open-weight distribution and low-friction API access. The DFlash speculator and the new enterprise-friendly licence are both moves that make sense if the goal is developer adoption rather than leaderboard position.

What It Means for Mid-Tier Labs

Poolside’s trajectory illustrates a structural shift that several mid-tier AI labs are navigating: the compute requirements for training at the frontier have grown faster than the venture capital available to fund them. Labs that raised large rounds in 2023 and 2024 at frontier ambitions are now making decisions about where to compete with smaller capital budgets.

Releasing capable open-weight coding models as free infrastructure — rather than as premium products competing against Claude Code or Cursor — is one answer to that constraint. Together AI’s $1.15 billion in bookings, announced the same week, represents the infrastructure side of the same trade: open-weight models are increasingly the production workload, and the labs that build developer access and serving infrastructure around them are capturing value even if they did not train the weights.

Laguna XS 2.1 does not carry published SWE-bench or Terminal-Bench numbers in the launch materials.