Discover.
Dispatches from the lab — product announcements, research notes, and field reports on what we're building and what we're learning.
Valar-27B: an AI scientist in 27 billion parameters
Introducing Valar-27B, a scientific reasoning model trained using Actuator to lead complete investigations across six scientific domains.
Introducing Marvin Atlas: search the world’s scientific literature
Introducing Marvin Atlas: 250M+ scientific records, with searchable text from 25M papers.
Kev: turn any LLM into a fast Jev-style decision model
Kev converts Qwen3 models into fast decision models that return choices and probabilities without generating text.
Marvin tops CompBioBench with 99/100, model roster updated
Marvin tops Genentech’s CompBioBench with 99/100, while Marvin Lite scores 95/100 in a single iteration.
How we cloned OpenAI's $230 Codex Micro in a weekend for under $40
AgentPad13 is an open-source take on the $230 Codex Micro and a test of whether an autonomous scientist can teach itself PCB routing.
When Scientific Output Outruns Scientific Judgment
AI is scaling research output but risks creating an unsustainable wave of slop.
Introducing the Iluvatar Open Research Initiative
Humanity faces more questions than scientific capacity. IORI is our effort to change that.
Introducing Actuator
A closed-loop control layer for model transformation: live monitoring and in-flight adjustment.
Divergent-Convergent Attention
We introduce Divergent-Convergent Attention (DCA), a transformer primitive that maintains parallel attention streams at different window sizes and reconciles them through learned periodic consensus.
Meet Marvin
We built Marvin because too much ML research time is still spent on preparation, context gathering, and logistics instead of the science itself.
Elastic Speculation
Elastic speculation delivers 30–50% lower latency and up to ~50% fewer speculative KV writes in our experiments, while preserving output quality.