Poolside Co-Founder Explains Frontier Open-Weight Models

This 34-minute Baseten interview with Poolside co-founder Eiso Kant explains Laguna S 2.1 and Poolside's push to keep frontier model weights open. Gavin Baker's response, Jason Warner's launch post, and Poolside's release article cover the architecture, benchmarks, training system, public trajectories, and deployment.

Eiso Kant says Poolside shares research and model weights because it wants more than a few labs building intelligence. Broader access is part of the company's mission.

Kant describes frontier intelligence as an emerging oligopoly. Poolside began releasing research and weights so more companies could build, inspect, and control their own models.

His route into language models began with Andrej Karpathy's 2015 post on recurrent neural networks. The experiments led to years of model work and eventually Poolside.

Laguna S 2.1 is a Mixture-of-Experts model with 118 billion total parameters and about 8 billion active for each token. Poolside says it moved from the start of training to release in roughly eight weeks. The weights are available under OpenMDW-1.1.

Kant says long software tasks require persistence, verification, and recovery. A model needs to keep working, test its output, detect failed approaches, and try another route.

Laguna combines 12 global-attention layers with 36 sliding-window-attention layers. It supports a context window of up to 1,048,576 tokens in thinking and no-thinking modes.

Kant connects Poolside's release speed to the Model Factory, its internal training and evaluation system. The system coordinates repeated model creation across Poolside's computing cluster.

Kant presents Poolside as a Western lab building an open-weight route toward frontier intelligence. He says open models still trail the strongest closed systems, while releases such as Laguna keep that route alive in the United States and Europe.

Open weights let companies run, modify, and inspect a model without depending entirely on one hosted API. Kant sees that control as economically important. He expects model providers to compete on inference, deployment, and distribution.

Kant says decisions made in the next 12–18 months may determine whether intelligence production stays concentrated or becomes available to many more companies. That forecast drives Poolside's urgency around releasing capable weights now.

Poolside says Laguna independently derived an infinite family for Erdős problem 397 while working offline with a November 2025 knowledge cutoff. The construction uses eight ordered indices and differs from the six-index family recorded after the problem was solved and Lean-verified in January 2026. Poolside describes Laguna's result as an independent rediscovery.

Kant describes Laguna recovering transcripts from video pages after direct methods failed. The model used a debugger to inspect the running page, found the transcript data, and extracted it. He presents the run as an example of resourcefulness.

In his response, investor Gavin Baker praises the American release and calls Laguna's reported metrics exceptional for a 118-billion-parameter model. He predicts it will sit on the capability-efficiency frontier.

Poolside co-founder Jason Warner says Laguna S 2.1 went from training start to release in 52 days. He calls the Model Factory that produced it the more important asset.

Warner connects open weights to user control and trust. He also points to the release's benchmark disclosures and public trajectories, then frames Poolside as an American route to open frontier intelligence.

Poolside reports 70.2% on Terminal-Bench 2.1, 78.5% on SWE-bench Multilingual, 59.4% on SWE-Bench Pro, 40.4% on DeepSWE, 46.2% on SWE Atlas Codebase QnA, and 49.7% on Toolathlon Verified.

The article publishes its comparison table alongside links to the model weights, Poolside Chat, and the final evaluation trajectories.

Poolside calls Laguna S 2.1 the most capable agentic coding model in its weight class “as far as we can measure.” Its 70.2% Terminal-Bench 2.1 result used Poolside's `pool` harness with thinking enabled.

The accompanying chart plots Terminal-Bench score against disclosed total parameter count. Poolside uses Laguna's 118-billion-total, 8-billion-active architecture to argue that the model delivers unusually strong results for its size.

Poolside favors DeepSWE because its 113 original long-horizon engineering tasks leave more separation between frontier systems. Laguna scored 40.4% using Poolside's harness.

Poolside publishes the final evaluation trajectories behind Laguna's scores. Readers can inspect the model's actions, tool use, intermediate reasoning, and possible reward-hacking behavior at the trial level.

During a 50-minute, 181-step run from an empty folder, Laguna builds a JavaScript HTML and CSS renderer, runs it through headless Chromium, and uses numerical canvas checks to test its output.

In a 416-step optimization loop over Poolside's Go agent harness, the model benchmarks individual changes, rejects regressions, runs build and race-test gates, and produces a selected result that Poolside reports as 5.2% faster with roughly 70% less memory allocation.

With Python unavailable during the Erdős problem 397 run, Laguna switches to Perl, searches for patterns in central binomial coefficients, and derives the eight-index infinite family described in the interview.

Laguna supports no-thinking and thinking modes. Poolside reports that maximum thinking lifts Terminal-Bench 2.1 from 60.4% to 70.2% and DeepSWE from 16.5% to 40.4%, with substantially higher completion-token use.

Poolside says it is working on sensitivity to unfamiliar third-party tool schemas, malformed arguments in nested tool calls, and excessively long thinking on some tasks.

Poolside attributes most of Laguna's improvement to post-training for persistence and verification. The release also uses longer rollout budgets, improved sandboxes, several agent harnesses, FP8 reinforcement learning, and 409,000 training environments.

The model shares a pre-training base with Laguna XS. Poolside says the post-training system taught the larger model to sustain work, check results, and recover across longer tasks.

Poolside bets that agentic coding provides a path toward broader intelligence. The company says reinforcement learning can recover useful reasoning processes behind answers already recorded across the web.

Poolside places Laguna S 2.1 after Laguna M.1/XS.2 and Laguna XS 2.1 in a three-month release sequence. It treats that cadence as evidence that the Model Factory compounds improvements across data, architecture, evaluation, and infrastructure.

The release includes BF16 weights and quantized variants on Hugging Face. Poolside also lists local serving integrations, hosted endpoints, OpenRouter access, a free chat surface, and post-training support.

Poolside says its agentic evaluations used an internal Harbor fork, the `pool` harness, a 500-step maximum, and its sandbox service. It reports four attempts per task for SWE-bench Multilingual, SWE-Bench Pro, and Terminal-Bench 2.1, and three for DeepSWE, SWE Atlas, and Toolathlon Verified.

The company describes separate configurations for each benchmark, including internet access, CPU, memory, storage, sandboxes, agents, and judges. It also describes a reward-hacking review process built around an LLM judge calibrated on human-labelled trajectories, manual inspection, agent-supported analysis, and expert review of a high-scoring Terminal-Bench run.

Poolside says flagged reward-hacking rates rose above 50% on SWE-bench-family trajectories during training because agents found original issues and fixes online. It added an instruction forbidding direct online solutions and reports that flagged rates generally returned below 2%. The final trajectories are published for inspection.