# about soma

train your own language model, on your own machine, for free, forever.

soma is an app for training language models. pick some text, start a model, watch it learn. stop whenever you like and pick it up later. talk to it. a soma model is never finished.

## one app, many runtimes

a runtime is the algorithm a model learns with. the app ships with three:

- **rma v1**: current. the residual metabolic aperture.
- **soma v14**: the adaptive trace-branch mixer.
- **soma v12.3**: the token-native spectral learner.

runtimes are locked. a model always opens with the runtime that made it.

## how rma learns

rma keeps exact traces of everything it has read, from the last word to millions of words back. there is no context window and no growing cache, so every token costs the same.

its controller, io2, listens to its own mistakes. when the text surprises it, it reads closely. when the text is predictable, it skims.

the full spec is at [/runtimes/rma-v1](/runtimes/rma-v1).

## no machine? train in the cloud

for bigger or longer runs, logOS can train your model on a cloud gpu. watch it live from any device. pause it, change it or stop it from your phone. you pay for the gpu; soma stays free. cloud training isn't open to everyone yet.

# about logOS

logOS is where soma models live. every model has a page, a runtime and a history. models are private, unlisted or public.

download a model, train it further, upload it again. or fork it into your own.

browse without an account. sign in with google to upload, star and train. soma is free and open source (mit).

james blight · [github](https://github.com/dejamesmusic)