Modelpublic

Akicou/Qwen3.8-Flash-Next-REAM-60Pct-GGUF

sourceHugging Faceotherupdated 28d agoView on Hugging Face

4likes861downloads

Model Card

Qwen3.8-Flash-Next-REAM-60Pct-GGUF

GGUF quantizations of Akicou/Qwen3.8-Flash-Next-REAM-60Pct, the REAM-compressed (Merged) version of Qwen/Qwen3.8-Flash-Next.

What this is

REAM (Router Expert Activation Merging) pruned 40% of the routed experts in the original model, taking each layer from 512 down to 308 experts. The compressed checkpoint was then converted to GGUF with `ggml-org/llama.cpp` (convert_hf_to_gguf.py, bf16) and quantized with llama-quantize. No importance matrix was used.

Files

File	Quant	Bits per weight	Size
`qwen4-ream-q8.gguf`	Q8_0	~8.5	~137 GB
`qwen4-ream-q4_k_s.gguf`	Q4KS	~4.5	~82 GB
`qwen4-ream-q4_k_m.gguf`	Q4KM	~5.5	~87 GB

The architecture is qwen4exp (hybrid linear attention + Qwen Sparse Attention MoE), 48 layers, 308 routed experts per layer.

Usage

bash

llama-cli -m qwen4-ream-q4_k_m.gguf -p "Explain reinforcement learning." -n 256

Notes

—Experimental release, not benchmarked.
—The base model requires trust_remote_code=True. These GGUF files are for llama.cpp (and compatible runtimes), so remote code is not needed at load.
—Shared experts, attention, and n-gram embeddings are untouched; only routed experts were merged.