Poolside’s Model Factory, Laguna S, Open Models, and the Race to AGI

This is a Latent Space interview with Poolside co-founder and co-CEO Eiso Kant, supplemented by Poolside’s July 21, 2026 Laguna S 2.1 release article. It covers Poolside’s origin, Model Factory, Laguna models, reinforcement learning, systems engineering, open models, regulation, financing, and AI-native engineering teams.

Andrej Karpathy’s 2015 character-level RNN demonstrations convinced Kant that learned systems could write code. His earlier company, source{d}, spent years applying machine learning to code and publishing its work. It failed after raising roughly $12 million.

Kant thinks source{d} stopped before scaling the learning system far enough. He left the field, then returned after ChatGPT confirmed his original direction. Poolside would pursue code models at foundation-model scale.

Poolside began with an AGI-through-reinforcement-learning plan. Open releases came later, while the company could still establish that practice before approaching the frontier.

Kant separates open weights from open research. Weights let others run and modify a model. Research shares methods and lessons that others can reproduce or extend. He leaves three release questions unresolved: misuse, sustainable economics, and whether future capability should change the company’s position.

Poolside is an American company with a global, remote, engineering-heavy team. It built its own model stack. Early optimizer failures slowed the company, but Kant says fixing them created the technical intuition and resilience behind its later speed.

Model Factory is Poolside’s versioned system for code, immutable streamed data, orchestration, training, inference, and evaluation. It shortens the path from an idea to a trusted result, then carries the result into the next model.

Early runs were hard to reproduce. The current system supports controlled ablations and automated integration. Laguna XS.2 went from the start of pre-training to release in about five weeks. Laguna S 2.1 followed in the same rapid sequence.

Agents already handle operational work inside the system. Kant says experiment velocity compounds because the process improves with every model.

Kant describes model development as repeated engineering: improve data, extract more from compute, evaluate, and repeat. He cites Zhipu, founded in 2019, as a lab that persisted before the market rewarded the category.

Poolside uses Model Factory to let more engineers conduct effective model research. Kant expects a wider field of suppliers because demand, national interests, and customer choice work against a two-provider intelligence market.

Laguna S 2.1 is a sparse Mixture-of-Experts model with 118 billion total parameters and about eight billion active per token. Poolside presents it as evidence that persistent behavior changes what a smaller model can do.

The model keeps working, verifies intermediate results, backtracks, and tries another route after failure. Poolside’s long-horizon coding evaluations and case studies document that behavior.

Kant thinks smaller open models may become the economic choice for much ordinary knowledge work, especially when they run locally. He still expects frontier progress to require larger systems.

Post-training produced much of Laguna S’s behavior. Kant expects the borders between pre-training, mid-training, supervised fine-tuning, and reinforcement learning to blur.

He thinks current training leaves useful signal in web data. Curriculum may become continuous, while automated searches discover objectives that researchers now select manually.

Model Factory runs the repeated experiments, environments, and evaluations behind this work. Kant still believes a lab needs a differentiated research bet alongside these standard scaling improvements.

Poolside treats matrix math, communication, numerical precision, sparse activation, attention layout, and reinforcement-learning throughput as capability work.

Laguna S 2.1 activates a small fraction of its parameters per token and mixes global with sliding-window attention. Poolside used FP8 reinforcement learning and released BF16, FP8, INT4, NVFP4, GGUF, and MLX variants. Quantized forms can run on one NVIDIA DGX Spark.

Each release feeds usage and failures back into the next Model Factory run.

Poolside trains core capability in its native harness, then does limited work to improve transfer. The harness affects efficiency and access. It can also change observed behavior through tool definitions, formatting, and context management.

The release article names harness overfitting as a Laguna S 2.1 limitation. The model may follow a remembered Poolside tool interface on its first attempt instead of a slightly different third-party schema.

Coding is Poolside’s route toward broader intelligence because it combines long-horizon reasoning with deterministic feedback. Software lets an agent act, test, and revise. Kant expects interfaces to move toward minimal computing environments where models compose actions in code. Current tool calling remains useful.

Poolside is beginning vision work. Audio is not a near-term priority.

Kant currently directs scarce compute and people toward language and long-horizon reasoning. He finds DeepSeek-OCR’s compressed visual-token approach interesting, but judges modalities by compute efficiency. Language already stores a dense record of knowledge and reasoning.

Poolside’s next, larger Laguna M had begun a 39-day training run. Kant calls the final run anticlimactic. Its visible compute cost excludes the research, failed experiments, data, infrastructure, and operations that made it possible.

Poolside announced a $500 million raise in October 2024 for talent, compute, research, data, and go-to-market work. Investors had to believe model capability would keep advancing and that machine intelligence would support valuable economic and scientific work.

The company was first going to be called Snowball Apps. The founders discovered a trademark conflict during early partnership talks and moved a sensitive conversation to a restaurant called Poolside. They kept the name as a reminder to choose the harder independent path.

Kant does not support openness at every possible capability level. He sees a future point for reconsideration.

At current capability, he rejects blanket restrictions. Elected governments should set boundaries rather than individual companies. He separates ordinary misuse from uncertain catastrophic scenarios and supports public safety institutes, evaluations, and capability-specific controls.

Regulation also operates under international competition. Rules that bind one country while models advance elsewhere may miss their purpose. Kant prefers limits tied to demonstrated capability.

Chip supply and allocation set the outer boundary. New accelerators and networks change which models are feasible.

Inside post-training, rollout generation becomes the bottleneck. Models produce long trajectories, interact with environments, and wait for results. Kant says one Poolside RL stage could not be shortened merely by adding GPUs because its batch and task structure limited throughput.

Prompt processing and token generation use compute and memory differently. Kant expects specialized resources, mixed hardware, and stable lower precision to shorten RL calendar time.

Large runs expose stability and scaling problems. Smaller runs allow faster iteration. Fresh pre-training helps Poolside test which recipe changes transfer.

Distillation answers a different deployment question: how to transfer a larger model’s behavior into a cheaper one. Poolside also wants small models to work as independent research experiments.

Laguna M.1, XS.2, XS 2.1, and S 2.1 form the published sequence. Poolside reports that S 2.1 is substantially stronger than M.1 at about half the active running size.

Kant measures engineering from the identification of user value to delivery. Code volume, pull requests, and reviews are intermediate signals.

He also asks how much one person can accomplish and how many people sit between a customer and a builder. In hiring, he looks for agency already demonstrated through work and personal projects.

Leaders align high-agency people through a shared outcome and explicit boundaries. Poolside’s reinforcement-learning focus creates those lanes. Kant says fewer than 70 people built the model and fewer than 115 researchers and engineers contributed to the wider effort.

Poolside’s July 21 article gives the model details: 118 billion total parameters, eight billion active per token, thinking and no-thinking modes, and a context window up to one million tokens. The company reports less than nine weeks from training start to launch.

Its reported scores are:

• Terminal-Bench 2.1: 70.2%
• SWE-bench Multilingual: 78.5%
• SWE-Bench Pro public dataset: 59.4%
• DeepSWE: 40.4%
• SWE Atlas: 46.2%
• Toolathlon Verified: 49.7%

Poolside calls Laguna S 2.1 the most capable agentic coding model in its measurable weight class. Its table mixes vendor results, benchmark leaderboards, and third-party maxima. The DeepSWE score came from Poolside’s `pool` harness, while the official leaderboard uses mini-swe-agent.

Poolside publishes trajectories from its final evaluations. It uses an LLM judge calibrated on human-labelled trajectories, manual inspection, and agent-supported analysis to look for reward hacking.

The release includes three long runs. Laguna S 2.1 worked on a browser engine without vision by using text artifacts and tests. It optimized an evaluation harness that Poolside checked with tests and performance measurements. It also explored an Erdős problem and produced a candidate result with a published trajectory and declared validation status.

Poolside reports higher scores with larger completion-token budgets on several benchmarks. The model exposes thinking controls, and `pool` lets users set thought level per session.

Poolside lists three:

1. A remembered native tool interface can override a similar third-party schema on the first attempt. Poolside says rejection and retry usually correct it through in-context learning. 2. Nested JSON arguments can become invalid or incorrectly escaped inside the model’s XML-like tool-call format. 3. Thinking can continue for unexpectedly long sequences before progress, especially on competition mathematics. Poolside plans to add effort control and improve thinking efficiency.

Poolside links Laguna S 2.1’s persistence to larger supervised fine-tuning and reinforcement-learning programs, broader task distributions, longer rollout budgets, FP8 RL, and training-stability changes.

Its first research bet is agentic coding: coherent work over long periods in software environments. The second is to recover reasoning omitted from written web answers through reinforcement learning.

Pre-training began on 4,096 NVIDIA H200 GPUs on May 22, 2026. Poolside credits the release speed to reused data work, architecture ablations, evaluation infrastructure, and post-training advances.

Laguna S 2.1 launched on Hugging Face under OpenMDW-1.1 in several formats and official conversions. Poolside announced support through NVIDIA tooling, vLLM, SGLang, Ollama, atomic.chat, Baseten, OpenRouter, Vercel AI Gateway, Kilo, Hermes Agent, pi, OpenCode, OpenClaw, Cline, and `pool`.

Poolside used an internal Harbor fork, its `pool` harness, a 500-step maximum, sandboxed execution, and repeated attempts.

SWE-bench Multilingual, SWE-Bench Pro, and Terminal-Bench 2.1 used four attempts per task. DeepSWE and SWE Atlas used three. Toolathlon used three runs. Adapters, resource ceilings, judges, and internet conditions varied.

Internet access was generally enabled unless stated otherwise. Poolside says its classifier eventually flagged more than half of some SWE-bench-family trajectories because models found original solutions online. An instruction against direct online solutions generally pushed the flagged rate below 2%, with named exceptions. Poolside also used manual and agent-supported review.