Poolside's Laguna S 2.1 matches larger models with 118B parameters, 8-week cycle

Eiso Kant's journey to Poolside began in 2015, when he read Andrej Karpathy's article 'The Unreasonable Effectiveness of Recurrent Neural Nets.' Inspired by the examples of character-level language models that could generate text and code, Kant pivoted his startup overnight to focus on building language models for code. He founded Sourced, a fully open-source company, and spent four or five years exploring convolutional neural networks applied to code structure, attention mechanisms on LSTMs, and Transformer models—all before the broader industry recognized the value of scaling language models. By the end of 2019, after spending $12 million（約19億円） of investor money and assembling a team of 40 people, the venture had failed. Kant later reflected that the team was on the right track but did not understand the importance of scaling; it was not obvious at the time that simply scaling up was the answer. He stepped back from language models for two years, focusing on family.

When ChatGPT launched, Kant felt vindicated. The broader market finally recognized that language models—including those trained on code—could be genuinely useful. He began rebuilding Poolside around a thesis that had animated his earlier work: as more capable AI is built, it should be open and open source. Crucially, Kant distinguishes between open weights (releasing model parameters) and genuinely open research (publishing methods and findings in detail). Poolside embraced both. The company deliberately built a global research organization outside the Bay Area talent war, betting that great researchers could be found anywhere and that diversity of location and perspective would strengthen the work.

At the core of Poolside's competitive advantage is the 'Model Factory,' an end-to-end engineering system for rapidly training and improving models. Fewer than 70 researchers run approximately 10,000–20,000 experiments per month. The key innovations include streaming data directly into training (unlocking faster experimentation), immutable data and versioned code (enabling reproducibility and scientific rigor), and a feedback loop so tight that engineers can often evaluate a new checkpoint within its first 30 minutes. This architecture has shrunk model cycles from six months to five and eight weeks.

Laguna S, released on July 21, 2026, is a 118B total parameter Mixture-of-Experts model with 8B activated per token, a context window of up to 1M tokens, and both thinking and no-thinking modes. The model holds its own against competitors that are nearly 10 times larger. Kant attributes this performance not to raw intelligence but to persistence, verification, and backtracking—mechanisms that allow the model to double-check its work and recover from errors. The eight-week development cycle represents the final step in a compression process that redefined what a small team could accomplish in model development.

Kant's philosophy extends beyond engineering to the structure of the AI industry itself. He believes that the path to AGI runs through coding and long-horizon software tasks, areas where agents that write scripts and modify their own training pipelines will eventually surpass systems that merely choose from predefined tools. He is dismissive of MCP (Model Context Protocol) and traditional tool calls as 'stupid,' preferring minimal harnesses and maximum model freedom. Poolside is prioritizing vision capabilities next but does not expect to work on audio soon, reflecting a view that language remains the most compute-efficient modality for encoding knowledge and reasoning.

Poolside raised $500 million（約800億円）, a significant vote of confidence in its thesis that many independent foundation model companies can coexist. Kant argues that intelligence could become the world's most demanded and commoditized resource, and that unilateral AI safety does not work in a globally competitive environment. He warns that regulation could accidentally lock in an oligopoly of two or three companies, and that open models may eventually become too capable to release without restrictions—a tension the industry has yet to resolve. The company sees NVIDIA and TSMC as the true infrastructure choke points in the AI supply chain, and identifies reinforcement-learning wall-clock time as one of its biggest remaining bottlenecks. Rather than simply distilling larger models, Poolside trains from scratch, a choice that reflects confidence in its engineering systems and a belief that smaller models trained correctly can outperform larger ones distilled down.

Poolside's release of Laguna S 2.1 marks a turning point in the open-source model landscape at a moment when the industry is locked in debate over model ownership, supply-chain sovereignty, and whether frontier AI development must consolidate into the hands of a few well-capitalized companies. The company's journey—starting with Eiso Kant's 2015 bet on Andrej Karpathy's work on recurrent neural networks for code, through a $12 million（約19億円） failure when the market did not yet care about language models on code, and finally vindication after ChatGPT proved the thesis right—illustrates both the long conviction required and the timing risk in frontier research. The company's pivot from a initially closed stance at founding to an explicit embrace of open weights and open research reflects a deliberate strategic choice: Kant states he would rather live in a world with 100 foundation model companies than five, even if Poolside were one of the five. This philosophy underpins not only the release model but also the company's engineering culture.

The 'Model Factory' system Poolside has built represents a radical compression of the model development cycle. Running 10,000–20,000 experiments per month with fewer than 70 researchers, streaming data directly into training, and practicing immutable data and reproducible experimentation has unlocked a move from six-month cycles to eight weeks—or, in some cases, five weeks. This acceleration is not merely a matter of throwing more compute at the problem; Kant emphasizes that model building is ultimately 90% engineering, and that 95% of model building can be reduced to better data or compute efficiency. The ability to evaluate a new checkpoint within its first 30 minutes suggests a feedback loop that traditional academic labs or smaller teams cannot match.

The competitive challenge Laguna S 2.1 poses to much larger models points to a deeper insight: smaller models may contain far more capability than previously assumed. The company's focus on persistence, verification, and backtracking—rather than raw model intelligence—suggests that the path to useful AI may not require ever-larger models but rather smarter architectures and training regimens. Poolside's $500 million（約800億円） raise, completed while investors were still questioning whether AGI was real, reflects a bet that this engineering-first, efficiency-focused approach is viable and preferable to the capital-intensive scaling race that has defined the sector thus far.

No comments yet. Be the first to share your thoughts!