Qwen3.8-Flash-Next-GSQ-RCO-Q8_0-PLE-GGUF

A GGUF containing the per-layer token embedding table (per_layer_token_embd.weight, 51.2B parameters) of Qwen3.8-Flash-Next quantized to Q8_0 — a PLE (Per-Layer-Embedding) companion shard.

It is intended to be used together with the main repository ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF, which provides the GSQ-RCO non-uniform quantizations of the transformer weights plus the BF16 vision projector (mmproj). There, the per-layer embedding table ships as a second shard at IQ4_NL; use this shard instead when you want that lookup table held at Q8_0.

File naming / -hf support

llama.cpp pairs split GGUFs by filename, not by metadata. To make llama-server -hf (and anything else that relies on filename matching) work without manual renaming, the same shard is provided under one filename per quant variant of the companion repo:

Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-00002-of-00002.gguf
Qwen3.8-Flash-Next-GSQ-RCO-IQ2_XS-00002-of-00002.gguf
Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf
Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00002-of-00002.gguf
Qwen3.8-Flash-Next-GSQ-RCO-Q8_0-00002-of-00002.gguf   (generic name, for custom builds)

All of these files are byte-identical (same content hash) — download only the one matching the quant you are pairing it with. Place it next to shard 1 (...-00001-of-00002.gguf) from the companion repo and load shard 1 as usual.

No metadata editing is required: the split.no / split.count / split.tensors.count values are identical across all GSQ-RCO builds, and the shard contains exactly the single per_layer_token_embd.weight tensor that each build's shard 2 holds at IQ4_NL.