Poolside

Poolside is an American AI foundation-model company founded in April 2023 by Jason Warner, formerly GitHub’s CTO, and Eiso Kant, founder of source{d}. It develops models for agentic coding and long-horizon software work, treating software development as a path toward artificial general intelligence. Its Model Factory integrates data processing, training and experimentation; its public Laguna releases in 2026 paired models with tools for using them.[4][7][1]

On August 20, 2026, Eric Newcomer and Tom Dotan reported, citing a letter to investors, that Poolside had struck a $6 billion non-exclusive licensing deal with NVIDIA, alongside a separate $1 billion investment in Poolside at a $12 billion pre-money valuation. The valuation was distinct from the licensing payment and investment.[12]

Latent Space’s August 21 account reported that NVIDIA was licensing the Model Factory and hiring 109 Poolside employees. It reproduced the founders’ description of the arrangement as “not an acquisition and not an acquihire,” and their account of losing a planned 40,000-GB300 cluster after failing to close a $2 billion financing within a six-week window at the end of 2025. The founders argued that frontier-model development was constrained by physical data-center space and contracted compute as well as capital. They were “not ready to share the updated vision” for Poolside.[2]

Poolside’s coding focus had a predecessor in Kant’s earlier work. In his July 2026 interview, Kant recalled pivoting his startup after reading Andrej Karpathy’s 2015 article “The Unreasonable Effectiveness of Recurrent Neural Nets.” At source{d}, he pursued language models for code through the end of 2019, experimenting with recurrent networks, LSTMs, attention and Transformers. He described the effort as a failure that consumed $12 million of investors’ money; that expenditure belonged to source{d}, before Poolside.[1]

Warner and Kant first met in 2017 and founded Poolside in April 2023. Kant said Poolside was American from inception and initially chose to recruit researchers outside the Bay Area, anticipating competition for talent. It operated remotely, opened offices including Paris and London, and recruited across the United States and internationally; it subsequently also had Silicon Valley staff. Kant recalled the roughly five-person early team spending three weeks resolving unstable training in its own training codebase, although he could not remember the exact optimizer bug.[4][1]

On October 2, 2024, Poolside announced the closing of a $500 million Series B led by Bain Capital Ventures. Felicis Ventures and Redpoint Ventures participated after leading its $100 million Series A and $26 million seed round respectively; corporate participants included NVIDIA. Poolside said the financing had brought 10,000 NVIDIA GPUs online. Its initial commercial focus was an intelligent coding assistant for large enterprises and the public sector, with go-to-market led by former GitHub CRO Paul St. John.[4]

Poolside’s early research centered on Reinforcement Learning from Code Execution Feedback (RLCEF): models attempt solutions, execute code and use the resulting feedback to improve reasoning and coding. In its October 2024 announcement, the company said its models explored millions of tasks across 130,000 real-world codebases, receiving execution feedback for every attempted solution. This was training-task scale, rather than a count of customers or developers using the product.[3]

On December 4, 2024, Poolside announced a first-party AWS partnership. Enterprises could contract directly through AWS, apply existing AWS spending commitments and run Poolside within their AWS Virtual Private Cloud. The announcement named malibu and point as its foundation models, with Trainium or NVIDIA inference options. Poolside distinguished customer-local fine-tuning from training its foundational models: customer data would remain within the customer’s environment, with no data shared back to Poolside. It also supported air-gapped, on-premises and other cloud deployments.[9]

Poolside acquired London-based Fern Labs, announcing the transaction on November 18, 2025. Ash Edwards, Alex Goddijn and Taylor Young had founded Fern after leading complex deployments at Palantir. Its Bridge software orchestrated multi-agent systems, while forward-deployed research engineers embedded with customers. Poolside connected the acquisition to deploying agents inside private enterprise environments and translating customer problems into reinforcement-learning tasks. The announcement acknowledged that reliable, long-running multi-agent systems in high-stakes environments remained an open problem.[10]

The Model Factory grew from Poolside’s decision to involve distributed-systems engineers in research from the beginning. It connects raw-data ingestion, filtering and transformation with distributed training, post-training and evaluation. Kant described its organizing metric as the time from a researcher’s idea to an experimental result the team could trust and incorporate into the next model.[1]

A central mechanism is streaming data into training rather than waiting for an entire dataset to be packaged and copied across the cluster. Poolside’s Blender service configures mixtures of data sources, repetition and shuffling, allowing training to begin while more data is still materializing. Immutable data and versioned experiment code let researchers trace which tokens entered a run and reproduce earlier experiments. Kant credited Nikolai, an applied-research head who joined from Yandex, with strengthening scientific rigor as the platform developed.[1]

During the July 15, 2026 interview, Kant described fewer than 70 researchers and another 35 engineers, estimating roughly 10,000–20,000 experiments a month while noting that he had not checked the latest count. He described Poolside as operating a 10,000-H200 cluster at that time. Agents were increasingly writing code, launching jobs, evaluating results and modifying pipelines, while researchers still supplied ideas and helped debug. Kant credited applied-research heads Peng Ming and Nikolai and the wider team for the models.[1]

Poolside also published hardware-efficiency research with NVIDIA. A March 18, 2026 post by Poolside’s Arkadii Bessonov and Joe Rowell and NVIDIA’s Ming Huang, Dan Blanaru, Sagar Singh and Matthew Radzihovsky described moving MLP activations to Grace host memory and reloading them during backpropagation. Overlapping transfers with computation avoids the recomputation cost of activation checkpointing, provided transfers finish within the available compute window. The authors measured approximately 4.3% better end-to-end step time on eight Qwen3-30B-3A layers and projected 6–13% higher throughput at full 48-layer depth. Their projected 0.5% peak-memory overhead differed from the experiment’s roughly 3% peak allocated-memory overhead, caused by an in-flight activation buffer. The separate 22% actual-memory gap was attributed primarily to allocator fragmentation.[11]

Kant said he and Warner renewed discussions about open source in early 2026, then chose to release weights and research to encourage more foundation-model companies. He distinguished providing weights from sharing the knowledge needed to recreate a model. His rationale was explicit: “I rather live in a world that has 100 foundation model companies than a world that has five, even if I was one of the five.” He also acknowledged unresolved questions about the business model and when misuse risks might justify restricting future releases.[1]

Poolside’s first public model release, on April 28, 2026, paired Laguna M.1 with Laguna XS.2. M.1 had 225 billion total parameters and 23 billion active parameters and had been pre-trained in 2025. XS.2 was a newer generation with 33 billion total and 3 billion active parameters, released as open weights under Apache 2.0. Poolside said XS.2 went from the start of pre-training to fully post-trained release in five weeks and credited approximately 60 Applied Research contributors across infrastructure, architecture, data, pre-training and reinforcement learning.[7]

Two products entered preview alongside those models: pool, a terminal-based coding agent, and Shimmer, a cloud development environment for iterating on web apps, APIs and CLIs inside a virtual-machine sandbox with the Poolside agent installed. The models were offered through Poolside’s inference endpoint and OpenRouter, initially free for a limited time. A May 26 update increased the served context lengths for M.1 and XS.2 to 256,000 tokens; Laguna XS 2.1 followed on July 2.[7][8]

Laguna S 2.1 launched on July 21, 2026. It was a mixture-of-experts model with 118 billion total parameters, 8 billion activated per token and up to one million tokens of context. Pre-training began May 22 on 4,096 NVIDIA H200 GPUs, putting training-to-launch time under nine weeks. It scaled the XS family using the same pre-training data as XS 2.1, with training-code fixes and recipe changes. Post-training combined supervised fine-tuning partly using synthetic data with reinforcement learning; changes included FP8 RL, longer rollout budgets, improved sandboxes and rollouts across multiple harnesses.[8]

Poolside reported S 2.1 scores of 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE v1.1, with thinking enabled in its pool harness, averaging pass@1 over four and three attempts per task respectively. These evaluations used an internal Harbor fork, up to 500 steps and five- or six-hour timeouts, emphasizing capability rather than execution speed. Comparisons selected maximum publicly reported scores and mixed harnesses. Poolside disclosed task modifications and reward-hacking mitigation, released final evaluation trajectories, and acknowledged third-party tool-schema problems, invalid JSON in nested tool arguments and overthinking. At launch, thinking could be switched off or set to max, without intermediate effort controls.[8]

The company illustrated S 2.1’s persistence with concrete case studies. It reported a 50-minute, 181-step session that built an HTML/CSS rendering engine from an empty folder without human intervention. Lacking vision capabilities, the model validated output through headless Chromium and numerical screenshot comparisons. In another automated feedback loop, it replaced quadratic string concatenation with buffers, memoized materializations and preallocated slices in Poolside’s own harness. Poolside reported a 5.2% speed improvement and roughly 70% lower memory allocation, qualifying the benchmark as not a full production test and validating the final artifact with Go’s race detector and go vet.[8]

Poolside also reported that S 2.1 rediscovered a proof of Erdős problem #397 in a 68-minute offline session using Perl, explicitly acknowledging GPT-5.2 Pro’s earlier January 2026 solution. S 2.1 weights were released under OpenMDW-1.1. NVIDIA helped optimize inference, while vLLM, SGLang and Ollama supported serving and local inference. Hosted distribution included Baseten, OpenRouter and Vercel AI Gateway; on OpenRouter, the free endpoint offered 256,000-token context and a dedicated paid endpoint offered the full one-million-token window.[8]