---

base_model:

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

library_name: gguf

tags:

• gguf
• qwen
• qwen3.8

---

Qwen3.8-Flash-Next-REAM-60Pct-GGUF-55GB

This is a static GGUF quant of Akicou/Qwen3.8-Flash-Next-REAM-60Pct.

I made this quant to keep the model weights below 64 GB so it has a better chance of running on systems with about 64 GB of combined memory.

The final quantized GGUF is:

• 60.84 GB / 56.67 GiB
• 3.779 effective BPW

Extra memory is still needed for the KV cache, compute buffers, the operating system, and other programs.

This model is heavily compressed. Quality may not be great compared with BF16 because there are deliberate precision cuts to make the model smaller.

No imatrix or calibration dataset was used. This is a direct static quantization.

Size comparison

| Model | Size | BPW |

|---|---:|---:|

| Qwen3.8-Flash-Next BF16 GGUF | 354.03 GB / 329.72 GiB | 16-bit weights |

| Qwen3.8-Flash-Next-REAM-60Pct BF16 GGUF | 257.67 GB / 239.97 GiB | 16-bit weights |

| This static quant | 60.84 GB / 56.67 GiB | 3.779 effective BPW |

The repo name uses 55GB as the target label. The exact finished size is shown above.

Quantization

• PLE n-gram table: Q4_0
• Routed expert gate/up: Q2_K
• Routed expert down: Q4_0
• Token embedding: Q4_K
• Output head: Q4_K
• Shared expert gate/up: Q3_K
• Shared expert down: Q4_0
• Most other compatible matrices: Q3_K
• Small or unsupported tensors stay in BF16/F32 when required by llama.cpp

Q2_0 and IQ quants are not used.

The quant was made directly from the BF16 GGUF in the BF16 folder using llama.cpp commit cc231cb0da565440cf6a3e5b55dfeba477972cb6.

The current llama.cpp Qwen3.8 converter does not export the separate MTP draft head, so the BF16 GGUF numbers here refer to the llama.cpp text-model export.