ClyxBox

Inference runtime for Clyx causal language models.

pip install clyxbox==0.3.2

from clyxbox import ClyxPipeline

pipe = ClyxPipeline.from_pretrained(
    "syntropic-clx/Clyx_0.3-635.51M-BASE",
    device="cuda",
)
text = pipe("Once upon a time", max_new_tokens=128, stream=True)

When dtype is omitted, ClyxBox selects native BF16 on supported CUDA GPUs (for example, A100) and FP32 otherwise (including T4). Pass dtype="float16" or dtype="bfloat16" only when you have verified that precision on your GPU.

The default sampling settings are temperature=0.65, top_p=0.90, top_k=50, and repetition_penalty=1.12. Sampling is stochastic, so outputs can differ between runs. A BASE checkpoint continues text and may produce repetitive or off-topic continuations; these settings cannot turn it into an instruction-following assistant.

The package supports Hugging Face model repositories and Clyx inference bundles, including exports split across multiple Safetensors files. For a sharded export, keep model.safetensors.index.json and every referenced shard in the same repository; ClyxModel.from_pretrained reads the index and loads each shard automatically. A BASE checkpoint continues text; it is not automatically an instruction-following assistant.