Seph · v0.1 · Seph-SYour model. Your machine.
A language model built from scratch, trained on rented GPUs, and quantized to run on nothing but a CPU. No API, no vendor, no one else’s weights.
A modality-agnostic backbone
Seph’s text-only v0.1 is not "a text model to extend later" — it’s a backbone whose text-only configuration is mathematically identical to a standard decoder. Every seam vision and audio will eventually need is already load-bearing, so growing into a multimodal model never means a rewrite or an invalidated checkpoint.
Embedding bus
The backbone takes inputs_embeds, never input_ids — every future modality is a registration, not surgery.
CPU-first
Aggressive GQA, a 32k vocabulary, and Q5_K_M quantization keep the whole thing runnable without a GPU.
Multi-token prediction
Trained in from step one — MTP heads can’t be added after the fact, and they’re the biggest CPU decode win available.
Forward-compatible checkpoints
Every checkpoint you train today is designed to keep loading into whatever Seph becomes next.
Ready when you are.
Every conversation here runs against your own checkpoint, on your own infrastructure.
Open chat