Home 00 CLI 01 Neuralese 02 Latent Reasoning 02.1 MultiThink 03 Research 04 Models 05 About 06

About Cymela

Cymela is an independent AI research effort working on continuous latent reasoning, the MultiThink architecture, the Hyper research checkpoints, and Monarch CLI, a terminal coding agent.

The short version
  • Cymela is an independent, self-funded AI research effort, started in 2026 and operated from Serbia.
  • We research how language models can reason in continuous internal state rather than only in generated text.
  • We publish our checkpoints, our measurements, and the results that argue against us.
  • Cymela is sometimes written Cymela AI to distinguish it from Megisto cymela, an unrelated butterfly species that shares the name.

This page is the canonical description of who Cymela is. Everything on it is intended to be verifiable from the links in Section 04.

SECTION 01

What Cymela is

Cymela is an independent artificial-intelligence research project, founded in 2026 and operated by a private individual in Serbia. It has no institutional backing, no outside investment, and no parent organisation. Research is self-funded and compute comes from free and low-cost public tiers, which is a real constraint on what we can run and one we state plainly rather than work around in our claims.

The work divides into two halves that feed each other. The first is research into how a language model can carry reasoning forward as internal state instead of as generated text. The second is tooling that puts models to work, which is where the research meets something people actually run.

Cymela is not a laboratory in the institutional sense, and we do not describe it as one. It is a small research effort that publishes in public.

On the name: Cymela and Cymela AI refer to the same project, and both are us. We are unaffiliated with camelAI, CAMEL-AI and Caimera AI, separate organisations with similar-sounding names, no connection to this one. The official channels in Section 04 are the complete list; anything else claiming to be Cymela is not.

SECTION 02

What we build

Neuralese: continuous latent reasoning

Neuralese is our research into reasoning that never becomes words. Instead of emitting tokens to think, the model recycles its own hidden state for a number of steps before producing any output, with a learned halt head deciding how many. The open question is whether that internal trajectory carries the content of a specific question, or only the shape of an answer in general.

MultiThink: models running concurrently

MultiThink is our architecture work on two independently trained models running at the same time rather than taking turns, one generating while the other checks and corrects. It is in active research, and the page documents what is built and measured as well as what is not.

Monarch: the research checkpoints

Monarch is the name of our model line. Hyper v1, the first checkpoint, predates that name and keeps it. It is published under it. Hyper v1 is a 3.10B-parameter checkpoint: Qwen2.5-3B-Instruct extended with our own latent-reasoning modules, trained for 82,697 steps. Weights are published. It is released under the Qwen RESEARCH LICENSE AGREEMENT for research and evaluation only, not for commercial use, and it has had no safety training of any kind. Built with Qwen.

Monarch CLI: a terminal coding agent

Monarch CLI reads, edits and runs code on your machine and talks to whichever model provider you configure. You bring the provider and the API key. There is no Cymela account, no telemetry, and no Cymela server anywhere in the path. It is distributed on npm and requires Node.js 20 or newer.

Hyper is the CLI's default persona, a tone and style preset. It shapes how the agent writes, not what runs underneath. Hyper is the frontier-model persona; Monarch is the one written to sit on any model. Monarch is also the name of our model line, and the two senses are not the same thing: no Cymela model powers the CLI today, and choosing a persona does not change which model answers you.

SECTION 03

How we publish

We publish numbers we have measured and leave out the ones we have not. Where a figure is uncertain or two measurements disagree, the site says so instead of picking the better-looking one. Where a result argues against the direction we are taking, it goes in the research log with the same prominence as anything else.

This is not modesty for its own sake. A small research effort has exactly one asset that scale cannot buy, and that is being right about what it claims. Overstating a result costs more than the result was worth.

The clearest example is on the record: we discovered that 79,137 of an 82,697-step training run had executed at a latent depth of one, because of a configuration bug, and that only 2,311 steps of the released checkpoint carry genuine multi-step training. We published that number rather than the configured one, and published the checkpoint anyway with the limitation stated.

SECTION 04

Official channels

These are the only channels operated by Cymela. Anything else claiming to be Cymela is not us.

There is no Cymela package on PyPI. The cymela name is reserved there as a signpost that points back to npm; installing it does nothing.

SECTION 05

Contact

Correspondence of any kind, technical questions, licensing, press, or legal notices, goes to contact@cymela.com. For bugs and feature requests in the CLI, an issue on GitHub is better than email, because the answer is then useful to somebody else.

Cymela is operated by a private individual and does not publish personal identifying information about its operator. Verified ownership and legal requests are handled through the address above, as set out in our Terms of Use.