downl

soma runs anywhere python does. pick the app for your platform, or grab the script bundle if you want the source.

see how it works →

apps

soma · mac app

183.2 MBreleased 2026-08-01

a minimal desktop app with four main spaces.

chat

load a checkpoint and talk to it locally — by text or by voice. prompts can optionally update the model as you chat, so the conversation becomes part of its ongoing training rather than a disposable session. the mac app can record one utterance through apple's on-device speech recognition and read replies through the macos system voice, with no network fallback. controls include response max, temperature, online learning, pass prompt, and checkpoint load/unload.

train

continue training a model from local text files or live streams. drop files into the app's data folder, choose a checkpoint, set the training shape and controller, then watch loss, accuracy, throughput, decimation, saves, and sample "dreams" as the model learns. training can be stopped with or without saving.

stream

run small live input scripts that feed soma an ongoing text stream. streams are treated like training sources, useful for experiments where the model keeps adapting to changing input.

logOS

browse shared checkpoints from logossoma.com, inspect their metadata, download them into soma, and continue them locally.

under the hood, soma keeps a local app folder for checkpoints, data, streams, logs, and chat history. the interface is intentionally plain — black and white, sf mono, direct controls, no cloud account required for local use. the app is built around the idea that a model should be something you can keep, train, interrupt, resume, talk to, and share.

download soma.dmgdrag soma to applications · open it · drop your corpus into soma's data folder, then hit train

script bundle

soma v12.3.17

released 2026-08-23

not on mac? soma v12.3.17 as a python script bundle. token-native continual spectral learner — a fixed 8k sentencepiece dictionary (soma-8k-v1) feeds a token trace bank whose retained identity is projected through the current learned dictionary each batch; a serial budget-residual mlp of depth D composes over the projected bandpass present with tied io/output dictionary and residual gates initialised at 0.1; row magnitude is governed by radial plasticity compression rather than hard row clipping, so rows lose outward growth as they approach capacity while inward and directional learning stay open; io2 drives plasticity, decimation gain and depth saturation from the spectral residual. zip containing the v12.3.17 runtime (soma_v12_3.py), the checked-in soma-8k-v1 tokenizer, a launcher shell wrapper, the logos bridge, the training loop and worker, stream sources, corpus builders, the current v12.3.17 spec, a cloud trainer, and the local speech helpers (soma_voice.py + macos_soma_speech.swift + soma-speech binary) so the mac app can talk and listen fully on-device. for linux, windows, cloud gpus, or anyone who wants the source.

three lines to a running soma

unzip soma.zip
cd soma-v12.3.17-script-bundle
./soma                # mac / linux — bootstraps a venv on first run

windows or any-platform manual

pip install -r requirements.txt
python soma_v12_3.py

bundle (.zip · 271.5 KB)soma v12.3.17 (.py)spec (.md)

reading

paper

time is all you need: temporal translation and the credit assignment problem · aaai 2026.

paper (.pdf)

downl_