Inside Poolside's Model Factory: 20,000 AI experiments a month, a coding model every 8 weeks

Poolside runs somewhere between 10,000 and 20,000 training experiments every month and ships a new coding model roughly every eight weeks. That cadence is the point. The lab has built an engineering factory around it, and on a recent Latent Space podcast co-founder Eiso Kant argued that the factory itself, not raw model size, is now the most important variable in competitive AI.

Poolside releases its models as open weights, the same category as Meta's Llama, Mistral, and the newly open-weights Thinking Machines' Inkling. The company's most recent ship, Laguna S 2.1, is a 118-billion-parameter mixture-of-experts model that activates about 8 billion parameters at inference and runs in a 1-million-token context window. The combination of a relatively small active footprint and a long context is Poolside's wager that code generation does not need frontier-scale compute to be useful.

Kant has been at this since 2015, when he started building code language models after reading Andrej Karpathy's early work on recurrent neural networks. He and his co-founder spent roughly four years and about $12 million before the launch of ChatGPT made code-LMs a market. "ChatGPT was vindication," he told Latent Space. The path since has been less about chasing a single flagship model and more about building the apparatus that produces them.

That apparatus is the Model Factory. Pre-training to release takes about eight weeks. Inside that window, every experiment is reproducible, training data streams continuously into the cluster, and compute runs at low precision to keep the marginal cost of each experiment tractable. The 10,000-to-20,000-experiment monthly volume is what makes the eight-week cycle possible: enough shots on goal to find a recipe that holds up across evals and downstream code tasks.

Kant describes a second, more novel feature of the factory: the same agents that write code for Poolside's customers are starting to write code that modifies the training pipeline itself. Agents launch jobs, evaluate results, and edit the recipes used to build the next model. If the loop holds, the model is no longer the end product of a research team; it is one output of a software system that improves itself each cycle. Poolside's framing, in a follow-up X post and on the founder's account, is that this loop is what an open-weights lab scales on when it cannot outspend the closed frontier.

The ideology runs through the cadence. Kant told Latent Space he would prefer 100 viable foundation-model companies to a five-lab oligopoly, and he draws a sharp line between releasing open weights and running genuinely open research. The distinction matters because Thinking Machines' Inkling release this month, an Apache 2.0 open-weights 975-billion-parameter model, sharpened exactly that question: a downloadable model is not the same as a research process a competitor can replicate. Poolside has raised about $500 million; the bet is that engineering cadence and a self-modifying pipeline are the moat, not parameter count.

Two questions hang over the factory. First, whether the eight-week clock survives contact with frontier-scale work. Laguna S 2.1 is competitive on coding evals against much larger models, but the comparison set is mostly other open-weights releases; independent head-to-head evals against the current closed frontier are thin. Second, whether the agent-modified-pipeline claim is durable. The current loop is mostly bounded: agents tune recipes and run jobs, but the high-level research direction is still set by humans. Crossing that line is the difference between an automated research assistant and an automated research lab, and Poolside is closer to the former.

The next release lands in about two months, on the same eight-week clock. Whether it closes the gap to the closed frontier, or merely holds the line, is the watch item.