Base Labs

Base Labs is a research organization spun out of AI infrastructure company Baseten, focused on post-training and training science for open-weight language models. Its stated agenda covers continual learning, the science of reinforcement learning, model memory in the KV cache, specialization, reasoning, and serving, with early work already published on topics like neural KV cache compaction, on-policy versus dense training, and whether language models can learn facts continually in their weights. The audience is researchers, engineers, and open-source model builders who want a counterweight to closed frontier labs. The lab operates on four public commitments: publishing without exception, including negative results, in plain English, encouraging skepticism as part of the scientific method, starting from concrete falsifiable problems, and refusing to sell a silver bullet before rigorous evaluation. Alongside the research group, Base Labs runs the BaseHub Data Foundry, a dedicated effort to produce high-quality RL environments for the open-source community, and a three-month residency in San Francisco called the Fellowship, which pairs researchers with a senior mentor and full compute access to produce one publishable result. Compensation is listed at $15,000 per month plus a compute budget, with cohorts of three fellows. This launch matters because it formalizes something the open ecosystem has been missing, a research group with production-grade compute and inference infrastructure sitting directly behind it. Baseten already serves open models at scale for enterprise customers, and Base Labs uses that position to run large experiments and release the recipes rather than keep them internal. For founders and operators watching the open versus closed model debate, it is a concrete bet that a healthy frontier depends on shared science, shared data pipelines, and shared training methodology.