Poolside AI has published a detailed technical report and released a new model, Laguna S (Laguna S 2.1). Co-founder Eiso Kant describes the company’s approach as a “Model Factory”: an engineering-led, end-to-end process that moves an idea from pre-training through reproducible experiments to release, with a focus on shortening calendar time between idea and trustworthy result.
What is the Model Factory?
Poolside frames model building as mostly an engineering problem: data pipelines, distributed training and operational systems. Key elements of the Model Factory include:
- streaming training data (instead of packaging and copying static datasets to clusters),
- an immutable data layer and versioned code for full reproducibility,
- automated agents that write code, launch jobs, evaluate checkpoints and modify future training pipelines,
- a high experiment velocity: Poolside says fewer than 70 researchers (and additional engineers) run roughly 10,000–20,000 experiments per month.
A practical result: Poolside reports that some models shifted from six-month cycles to five- or eight-week cycles (e.g., Laguna XS launched in five weeks; Laguna S took eight weeks from the start of training to launch). Kant also highlighted that reliable infra has meant virtually zero on-call incidents during the model runs for much of the year.
Technical practices and experimental rigor
Poolside emphasises treating the entire pipeline as a first-class engineering problem. They stream data into training, which reduces time lost when datasets change, and expose data-mixing via configuration (Poolside’s “Blender” service). Immutable data plus experiments-as-code makes every run traceable down to token-level cursors, enabling rigorous ablations and reproducibility.
The team historically wrote a training stack from scratch and described debugging lessons (for example, an optimizer epsilon-related bug that affected stability and required corrective intuition). They treat these engineering investments as enabling scientific progress rather than just operational overhead.
Hardware and numeric formats: Laguna S used FP8 in training for parts of the pipeline. Poolside calls out MatMul and networking bottlenecks and RL wall-clock time as remaining bottlenecks, and is interested in lower-precision formats (NVFP4 / 4-bit / ternary) and mixed hardware approaches.
Laguna S: size, behaviour and performance
Technical highlights provided by Poolside:
- Laguna S: 118 billion total parameters, with 8 billion active (sparse) parameters,
- the model fits on a DGX Spark and Poolside reports 30–40 tokens/sec on a Spark in their measurements,
- model development cycles can be five to eight weeks from training start to public release,
- Poolside claims Laguna S outperforms some models that are multiple times larger on several benchmarks and real workloads (the report and conversation explicitly contrast Laguna S’s results with a recent Thinking Machines release that was nearly ten times larger).
Poolside researchers argue that much of Laguna S’s practical gains come from behavior changes achieved during post-training (persistence, verification, backtracking) rather than solely increased raw intelligence. That suggests smaller models may be made significantly more useful through behavioural post-training and RL-focused work.
Reinforcement learning and the training pipeline
Poolside believes RL will move earlier in training. They argue that simple next-token prediction is not extracting enough structure from web data, and that better curricula or objectives across pre-training could teach the model to reason earlier. Post-training RL to induce behaviours remains important, but Poolside is exploring ways to turn web-scale data into more effective curricula and to incorporate RL-style objectives earlier.
Open weights vs open research, and the argument for more competitors
Kant stresses the difference between releasing model weights and publishing genuinely open research: weights alone don’t let others reproduce the full pipeline (datasets, infra, recipes). Poolside released detailed technical material because they want more foundation-model companies to exist; Kant says he’d prefer a world with 100 foundation model firms rather than an oligopoly of five, and released research and weights to accelerate that outcome.
Poolside is legally an American company but deliberately built a distributed global research organisation — avoiding concentration in the Bay Area — to access talent worldwide and to avoid being a single-lab spinout.
Funding, regulation and ecosystem risks
Poolside has raised $500 million, according to the discussion. Kant argues government and democratic processes should play an active role in delineating risks and responses, warning that unilateral closures or blanket restrictions could inadvertently lock in oligopolies. He also highlights that hardware vendors (NVIDIA, TSMC etc.) and their supply dynamics strongly influence who can train frontier models.
Hiring and team impact
Poolside is hiring broadly across applied research, architecture, training, post-training, evaluations and engineering. Kant emphasises the high individual impact possible at a smaller, high-agency organisation: he notes that fewer than 70 researchers and roughly 35 engineers contributed to the released models, and that in such teams a single individual’s work can change the company’s trajectory.
Why this matters
Poolside’s experience shows that accelerated, engineering-first model development—streaming data, immutable datasets, reproducible experiments and automated agents—can compress the calendar time to new model releases and enable aggressive experiment rates. Their claim that behavioural improvements and verification can unlock much more capability in relatively small models is significant for debates about model commoditization and open-source strategies. At the same time, the company highlights hardware constraints, RL wall-clock bottlenecks, and regulatory trade-offs as key factors shaping who builds future foundation models.
Key facts and numbers
- Laguna S: 118B total parameters; 8B active (sparse) parameters.
- Model cycles reported: 5–8 weeks from training start to launch for certain small/medium models.
- Experiment scale: roughly 10,000–20,000 experiments per month with under 70 researchers and ~35 engineers.
- Prior spend mentioned by Kant (historical): ~$12 million spent building code-focused language models at an earlier startup.
- Poolside fundraise mentioned: $500 million.
Takeaways
Poolside’s Laguna S and the accompanying technical write-up foreground engineering practices as the decisive lever for faster iteration and more reproducible model science. They also make a case that smaller, well-behaved models can do far more practical knowledge work than previously assumed, and that open research and a competitive ecosystem are strategically important for the future of foundation models.
