D1-3B FP8 for Radiance
An unofficial FP8-quantized conversion of LiquidAI/d1-3B, distributed as d1-3b-fp8.rad for the Radiance runtime.
Modified artifact: the upstream weights have been converted and quantized into Radiance's .rad format. This is not a Transformers, GGUF, or safetensors checkpoint. Kernel optimizations belong to the runtime and are not bundled in this model file. See d1-3b-fp8.rad.NOTICE.
Usage
With a compatible Radiance installation and its plugins available:
radiance --model /path/to/d1-3b-fp8.rad --tp 1 --port 8000 --max-model-len 16384
For a build-tree installation, set RADIANCE_HOME to its build/radiance_home directory. This artifact was tested locally on an AMD Radeon RX 9060 XT with ROCm; compatibility with other hardware has not been verified here.
License and attribution
Original model: Liquid AI, Inc. Redistribution is under the LFM Open License v1.0, included verbatim in LICENSE. This license includes commercial-use revenue restrictions; consult its full terms. Original model documentation and attribution are preserved in UPSTREAM_README.md. This repository is not affiliated with or endorsed by Liquid AI.
SHA256SUMS contains the model file's SHA-256 checksum.