• Manifesto
  • Agenda
  • Articles
  • Fellowship

Working to advance and democratize open-source intelligence.

The Lab

  • Manifesto
  • Agenda
  • Articles
  • Fellowship

Stay in the know

© 2026 Base Labs

A research lab by Baseten

We're a research lab by Baseten working to advance and democratize open-source intelligence.

Our work spans model specialization, learning, memory, reasoning, and serving. We do all this work in the open, publishing our experiments, data, results in full.

Manifesto
Agenda
Articles
Fellowship

Recent research from the lab.

01

Can a Language Model Learn Facts Continually in Its Weights?

Research·JUL 2026
Charles
02

Post-Training Science for Supervised Fine-Tuning

Research·JUN 2026
Charles
Mudith
Harry
03

Still: Amortized KV Cache Compaction in a Single Forward Pass

Research·JUN 2026
Charles
alex
Harry
Mudith
max kirkby
04

Towards infinite context windows: neural KV cache compaction

Research·APR 2026
Charles
alex
Harry
05

Dense, on-policy, or both?

Research·MAR 2026
max kirkby
Charles
View all research

We are hiring across the lab.

We're built around researchers and engineers who have spent real time inside production systems, and PhDs working on problems that don't yet have a name. If that's you - come find us

Research Scientist, Post-Training

SF

Remote

Research Engineer, Post-Training

SF

Remote

View all jobs

Three months. One problem. Full support.

The Fellowship is a structured 3-month residency for researchers and engineers who want to spend uninterrupted time on a single hard problem in post-training.

Join the Fellowship program

3 months in San Francisco

Next cohort September 2026

You'll be paired with a senior researcher, given full access to the lab's compute and infrastructure, and expected to produce one publishable result by the end. Many fellows have continued as full-time members of the lab.

Compensation

$15,000 stipend for 12 weeks

Cohort size

3 fellows per quarter

Open to

PhDs, engineers and self-taught researchers

Deadline

Rolling, reviewed monthly

Apply nowLearn more

Four commitments on how we work.

I

We publish without exception.

We believe that AI research should be done in the open, without obfuscation and without secrets. We publish our results, including the negative ones. We also write in plain English so people can actually read our work.

II

We encourage skepticism.

We believe criticism is a core part of the scientific method and right now our field could do with more skepticism and debate. We hope that by making our results, methods and data public whenever possible, we can have productive conversations in the open.

III

We start with a concrete problem.

We pick problems that lead to clear, falsifiable results, and goalposts that we cannot move until we cross them.

IV

We will not sell or believe in a silver bullet until rigorously evaluated.

We don’t sell the snake oil of an “understanding of deep learning” or a “solution to continual learning”, we frame every scientific finding in its proper context.

We believe that AI research should be done in the open, without obfuscation and without secrets. We publish our results, including the negative ones. We also write in plain English so people can actually read our work.

We believe criticism is a core part of the scientific method and right now our field could do with more skepticism and debate. We hope that by making our results, methods and data public whenever possible, we can have productive conversations in the open.

We pick problems that lead to clear, falsifiable results, and goalposts that we cannot move until we cross them.

We don’t sell the snake oil of an “understanding of deep learning” or a “solution to continual learning”, we frame every scientific finding in its proper context.