FoxlightAI/qwen3-8-27b-fp8-mtp

FoxlightAI/qwen3-8-27b-fp8-mtp

This repository contains an MTP speculative-decoding sidecar derived from `Qwen/Qwen3.8-27B-FP8`. It is not a standalone model — it provides the multi-token-prediction heads used by Skulk to speculatively decode for the target model `Qwen/Qwen3.8-27B-FP8`. The heads ship at full precision (bf16, unquantized) — they are the speculative drafter, where precision drives draft acceptance — so there is one sidecar per base model and it serves every quantization of the target.

Provenance

Usage

Skulk loads this sidecar (mtp.safetensors) alongside the target model to enable MTP speculative decoding. It is referenced from the Skulk Weights Publisher catalog and fetched automatically by the Skulk shard downloader; it is not intended to be loaded standalone.

License

This artifact is derived from `Qwen/Qwen3.8-27B-FP8` and is published under that model's original license (apache-2.0), preserved unchanged. Refer to the source model's card for the full terms.