---

base_model: deepseek-ai/DeepSeek-V4.1-Flash

library_name: gguf

license: mit

tags:

• ranma

---

DeepSeek-V4.1-Flash RANMA Engram GGUF

> [!WARNING]

> Requires ranma.cpp ranma_20261013 or newer (github.com/saturnsky/ranma.cpp).

> This is not a model on its own: it holds only the engram tables that the DeepSeek-V4.1-Flash GGUF files of ranma.cpp need.

ranma.cpp is a llama.cpp fork developed with RDNA4 as its main target. It runs MoE models on a

three-tier expert cache: VRAM, host RAM and SSD. It supports EXL3 quantization to keep memory traffic as low as possible at a given quality.

What this is

The two engram tables of deepseek-ai/DeepSeek-V4.1-Flash (layers 1 and 14,

about 384 million rows each), taken from the original FP8 checkpoint without changing a value: each row is the FP8 E4M3 values and

their E8M0 scales, stored as MXFP8. Every quantization of the model uses the same tables, so they live here once and the model

repositories link to this one.

| Folder | Content | Size | Files |

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

| engram-MXFP8 | both tables, MXFP8 | 202.8 GB (188.8 GiB) | 7 |

manifest.json lists the size and SHA-256 of every file.

Used by

• SaturnHeaven/DeepSeek-V4.1-Flash-EXL3-3.0bpw-RANMA-EXL3-GGUF (3.0 bpw, Mia-AiLab)
• SaturnHeaven/DeepSeek-V4.1-Flash-UNCENSORED-EXL3-2.9bpw-RANMA-EXL3-GGUF (2.9 bpw, dealignai)

Download the folder once and give ranma its first file, --engram-tables .../DeepSeek-V4.1-Flash-engram-MXFP8-00001-of-00007.gguf;

the other files are found by name. At load the tables are checked against the model (table names, rows, row width) and

every piece is checked at both ends. How the set is found and checked:

engram-tables.md.

License

Same as the original model: MIT, see the original repository.