SpeedAI Lab is the research organization behind SpeedAI, a platform that helps researchers search, read, write, and reason over scientific literature.

Our research is organized around three directions: the architectures that make large language models cheaper and more capable, the agents built on top of them, and the scientific workflows they can automate. We prefer work that ships — open code, released checkpoints, and reproducible evaluation setups — so each direction below is paired with what we have released.

Making attention cheap over long contexts, without giving up retrieval quality.

• Gated linear attention and state-space hybrids (GDN, KDA, GDN2, Mamba-2, RWKV-7) as alternatives to full attention.
• Memory design — how recurrent state is initialized and allocated across heads and layers.
• Reproducible training and evaluation: one entry point for many architecture × method combinations, with published checkpoints, samples, and predictions.

Turning models into systems that can carry out multi-step work rather than answer single questions.

• Long-horizon task execution, planning, and tool use.
• Retrieval and memory for agents that operate over large document collections.
• Evaluation of agent reliability under real, noisy research workloads.

Applying the above to how research itself gets done, and to the scientific models that do the work.

• Scientific models — domain foundation models trained on scientific data: molecular structures, materials, biological sequences, and experimental measurements. We study pretraining objectives, data curation and representation, scaling behavior, and evaluation against domain-specific tasks rather than generic benchmarks.
• Scientific reasoning — how models acquire and use domain knowledge, and where they still fail on quantitative, mechanistic, and experimental reasoning.
• Literature understanding, synthesis, and writing assistance.
• Agentic research tooling for search, reading, and data analysis.
• Benchmarks that measure usefulness to researchers instead of benchmark-only scores.

What is ready to use today comes mainly from the architecture line:

For collaborations, research internships, or questions about the code, write to us — we are always glad to talk with people working on the same problems.

• Website: speedai.com
• Email: team@speedai.com