Specter: Fast MoE Inference via Quantization-Enabled Multi-Step Speculative Offloading
Paper • 2511.14102 • Published • 22
Specter DeepSeek-V2-Lite routed-expert GPTQ INT4 draft
Prepared routed-expert GPTQ INT4 draft checkpoint for Specter, accepted to ASPLOS 2027.
This archive mirrors the verified checkpoint published in the Specter GitHub release. It uses the Specter checkpoint layout and runtime. The draft quantizes routed experts to GPTQ INT4; other components retain floating-point precision.
Download and use
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="wwf606/Specter-DeepSeek-V2-Lite-draft", filename="specter-ae-dsv2-draft.tar.gz", local_dir="downloads")
Extract the archive and set the corresponding checkpoint path in the Specter configuration. Follow the source repository instructions for environment setup and inference.
SHA-256: a6f217f04197ed931cc0fa7833cdad8d091c528a8687112598d87d7bea169e29
The checkpoint licenses and notices are included inside the archive. Retain them when redistributing. Source code, dependency wheels, and prepared datasets are available in the archived artifact.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support