train your own ai, on your own machine, for free, for as long as you like.

soma is a mac app for training a language model locally and continuously. choose text, start a model, watch it learn, stop whenever you like, then resume the same checkpoint on the same or a different corpus. chat can continue updating the model. the artifact is never considered finished.

v14 uses a fixed 1,024-token sentencepiece dictionary (soma-1k-v1) so the network can learn composition rather than relearn spelling. an exact nonlearned float64 trace bank of shape (1,024 × 48) remembers which symbols occurred at which timescales; the temporal lens is an adaptive lattice on log time that io2 may conservatively redistribute. three shared scale blocks organise the temporal field, twelve fixed gaussian summaries yield a 1,536-wide context, and four residual swiglu blocks compose. the 384-dim branch reads through the same tied dictionary used at input, with the empirical log-prior added. time is held by the trace bank plus the lattice, not a context window or kv cache.

inference and training have constant cost per token. chat with online learning off to run at inference speed; leave it on and every conversation becomes part of the model. on the mac app, chat also works by voice — recognition uses apple's on-device path and playback uses the macos system voice, so the whole loop stays on your machine. view the technical specification for full detail.

logOS is where users share soma checkpoints, saved as single .pt files with an adjacent .meta.json. download one, continue training, and reupload to share your progress.

each checkpoint has a deterministic id — a hash of its state, unfakeable. each upload declares which checkpoint it was built on top of.

no login required to browse. sign in to upload a checkpoint. files up to 16gb. runtime is always free.