Independent research into continuous latent reasoning, MultiThink architecture, and a terminal coding agent.
Cymela's core thesis is continuous latent reasoning: instead of a model thinking only by generating text token-by-token, we're researching how it can carry reasoning forward as a continuous internal state — closer to how the underlying computation actually happens, before it ever gets forced into words.
Our current research checkpoint, Hyper, is Qwen2.5-3B-Instruct extended with our own experimental latent-reasoning modules — built with Qwen, and released under the same research-only license. It's early, and the behavior is still inconsistent: the model now thinks in latent space and that thinking helps, but it isn't yet specific to the question you asked. That gap is the work.
Alongside the model research, we build and ship Cymela CLI, a terminal coding agent you can install and use right now. One is a research bet; the other already works.
Our core research thesis: models that carry reasoning forward as continuous internal state instead of only through generated text. One checkpoint trained, weights published, results mixed.
Read about Neuralese →
A terminal coding agent, shipped: file editing, search, git, and shell tools, working across multiple model providers today.
Install the CLI →
An early research direction we call MultiThink: two models working at the same time rather than in turn, one generating while the other checks and corrects. Right now this is exploratory, not a shipped system.
See where it stands →
We train across whatever compute we can get: cloud TPU access and personal GPU hardware. Resourceful, not enterprise-scale — yet.
Explore the roadmap →