Qwen-3.5-28B-A3B-REAP — GGUF Q4_K_M

GGUF quantization of 0xSero/Qwen-3.5-28B-A3B-REAP, a pruned variant of Qwen/Qwen3.5-35B-A3B using the REAP (Refined Expert Activation Pruning) method.

Available Files

File	Quant	Size	BPW	Description
`Qwen-3.5-28B-A3B-REAP-BF16.gguf`	BF16	53 GB	16.01	Full precision, for re-quantization
`Qwen-3.5-28B-A3B-REAP-Q4_K_M.gguf`	Q4_K_M	17 GB	4.89	4-bit medium, recommended for most users
`Qwen3.5-28B-A3B-REAP-IQ3_XXS.gguf`	IQ3_XXS	11 GB	3.16	3-bit imatrix, smallest size with good quality

RAM / VRAM Estimation

Each quantization requires roughly model file size + 1.5–2 GB overhead for KV cache and runtime buffers (at default 4K context). Larger contexts will increase memory usage.

File	Quant	File Size	Est. RAM Usage
`Qwen-3.5-28B-A3B-REAP-BF16.gguf`	BF16	53 GB	~55 GB
`Qwen-3.5-28B-A3B-REAP-Q4_K_M.gguf`	Q4_K_M	17 GB	~19 GB
`Qwen3.5-28B-A3B-REAP-IQ3_XXS.gguf`	IQ3_XXS	11 GB	~13 GB

Which Quant Fits Your RAM?

System RAM	IQ3_XXS (~13 GB)	Q4_K_M (~19 GB)	BF16 (~55 GB)
8 GB	❌	❌	❌
16 GB	✅	❌	❌
32 GB	✅	✅	❌
64 GB	✅	✅	✅

Tip: If running on GPU, the same estimates apply to VRAM. With -ngl 99 (full GPU offload), you need the above amounts in VRAM. Partial offload splits the load between VRAM and system RAM.

Model Details

Property	Value
Architecture	Qwen3.5 MoE (linear + full attention hybrid)
Parameters	28.24B total / 3B active
Experts	205 total / 8 active per token
Context Length	262,144 tokens
Original dtype	BF16
Quantization	Q4_K_M (4.89 BPW)
Quantization tool	llama.cpp b8565
License	Apache 2.0

Quantization Process

# 1. Convert BF16 SafeTensors → GGUF
python convert_hf_to_gguf.py 0xSero/Qwen-3.5-28B-A3B-REAP \
  --outfile Qwen-3.5-28B-A3B-REAP-BF16.gguf \
  --outtype bf16

# 2a. Quantize to Q4_K_M
llama-quantize Qwen-3.5-28B-A3B-REAP-BF16.gguf \
  Qwen-3.5-28B-A3B-REAP-Q4_K_M.gguf Q4_K_M

# 2b. Generate imatrix (wikitext-2, 128 chunks)
llama-imatrix -m Qwen-3.5-28B-A3B-REAP-BF16.gguf \
  -f wiki.test.raw -o imatrix.dat --chunks 128

# 2c. Quantize to IQ3_XXS with imatrix
llama-quantize --imatrix imatrix.dat \
  Qwen-3.5-28B-A3B-REAP-BF16.gguf \
  Qwen3.5-28B-A3B-REAP-IQ3_XXS.gguf IQ3_XXS

Usage

llama.cpp

llama-cli \
  -m Qwen-3.5-28B-A3B-REAP-Q4_K_M.gguf \
  -ngl 99 \
  -c 4096 \
  -p "Your prompt here"

llama-server (OpenAI-compatible API)

llama-server \
  -m Qwen-3.5-28B-A3B-REAP-Q4_K_M.gguf \
  -ngl 99 \
  -c 4096 \
  --port 8080

LM Studio / Jan / Ollama

Download the .gguf file and load it directly in your preferred local inference UI.

Hardware Requirements

Config	VRAM / RAM
Full GPU (recommended)	20+ GB VRAM
Hybrid CPU+GPU	12 GB VRAM + 16 GB RAM
CPU only	24+ GB RAM

About the Original Model

0xSero/Qwen-3.5-28B-A3B-REAP applies REAP expert pruning (arXiv:2510.13999) to reduce Qwen3.5-35B-A3B from 128 to fewer experts while preserving performance. The result is a 28B-parameter model with only 3B active parameters per forward pass.

License

Apache 2.0 — inherited from the original model.

barozp/Qwen-3.5-28B-A3B-REAP-GGUF