télosWing It Research ↗

telos-ml

Train, decode, and evaluate language models.

Autoregressive, masked diffusion, uniform-noise diffusion, and COROSred. Python and CLI, with PyTorch and MLX backends.

pip install telos-ml

# Train a tokenizer and prepare your dataset
telos dataprep \
  --corpus data/python/ --output data/python.bin \
  --train-tokenizer --tokenizer tokenizer.json --vocab-size 8192

# Train a COROSred model
telos train --paradigm corosred --params 50M \
  --tokens 2.5B --data data/python.bin --tokenizer tokenizer.json \
  --checkpoint-dir checkpoints/run --hardware auto

What’s in the library

Data & training

Tokenization, dataset preparation, model configuration, checkpointing, and training across MLX, CUDA, and TPU.

Models & decoding

Transformer implementations, causal generation, diffusion sampling, and COROSred draft–refine decoding.

Evaluation

Contextual probes, functional code tests, span infilling, language and tool-use suites, and hardware benchmarks.

Side note: Apple Foundation Models

On macOS 27, Apple’s fm CLI provides direct access:

fm respond "Explain binary search."

Telos also includes an optional telos.afm bridge and evaluation adapter.

Apple’s tools ↗