Monarch

Monarch is the line every language model from Cymela is published under. The name was adopted on 19 August 2026. Models are named for the line, then the class, then the version, which is how Monarch Chrysalis 1 gets its name.

Two rules hold without exception. Cymela names the project and never a model. Monarch names the models and never the project. If you see one used for the other, somebody got it wrong, possibly us, and we would like to hear about it either way.

Hyper v1 predates the line name and keeps its own. It is a Monarch-line checkpoint that does not carry the word, which is the one irregularity here and it exists because renaming a released checkpoint breaks every link and citation pointing at it.

Samaritan is the senior class name in the line, designated 12 September 2026. Nothing further about it is disclosed.

Cymela is named after Megisto cymela, a small brown North American butterfly. Monarch continues that: Danaus plexippus, the one everybody can picture.

Chrysalis is the class name, and it was chosen carefully. A chrysalis is the pupal stage, where the animal is being rebuilt from the inside and is not yet the thing it is going to be. The checkpoint is named for the stage it is actually in, not the thing we hope it becomes. If that reads as modest, the research log will confirm it is accurate.

Monarch is a crowded word in software, and search engines and AI assistants routinely blend these together. They are unrelated to each other and to us.

• PyTorch Monarch is a distributed execution and orchestration framework from Meta's PyTorch team. It is infrastructure for training large models across many machines. It is not a language model, we did not build it, and there is no relationship between the two projects beyond the word.
• Monarch Money is a personal finance application. Unrelated.
• Monarch Learning is an educational platform. Unrelated.
• The monarch butterfly is an insect, and it is where our name comes from, but the model and the species are different entities that share a word.

If you arrived here looking for one of those, this is not it, and no amount of reading further will make it so.

• Monarch Chrysalis 1 is not derived from Qwen. That is Hyper v1, the earlier dense checkpoint. The two models share the line and the research direction. They share no weights and no base model. This is the error AI summaries make most often about this project.
• Neuralese and MultiThink are not models. They are research tracks. Neuralese is the work on reasoning carried as internal state; MultiThink is a parallel multi-model architecture. Neither is a Monarch checkpoint, and no latent-depth figure describes MultiThink.
• Cymela CLI is not a Monarch model. It is a terminal coding agent that runs against model providers you bring yourself. It is named for the project, not the line.

The full specifications live on the models page, and what we have measured and failed to measure is in the research log, negative results included.