---

license: mit

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

base_model_relation: quantized

library_name: llama.cpp

pipeline_tag: text-generation

tags:

• gguf
• llama.cpp
• deepseek
• deepseek-v4.1
• quantized
• text-generation
• long-context

---

DeepSeek-V4.1-Flash GGUF

Community GGUF quantizations of deepseek-ai/DeepSeek-V4.1-Flash.

<div align="center" style="background-color:#f59e0b;color:#ffffff;padding:16px 20px;border-radius:10px;line-height:1.7;">

☕ If this GGUF made your day easier, a coffee would make mine.<br>

<a href="https://ko-fi.com/ngquocvinh" style="color:#ffffff;"><strong style="color:#ffffff;">Send a coffee ☕</strong></a><br>

I build and test these releases myself. Your coffee helps keep me going.<br>

Thank you for supporting this work.

</div>

Release status

This model card is ready for the release package. Quantized files will be added

only after conversion, GGUF loading, and text generation have passed validation.

No GGUF artifact is published in this initial card revision.

The planned text-only ladder is:

| Priority | Planned file | Status |

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

| 1 | DeepSeek-V4.1-Flash-Q1_0.gguf | pending conversion and validation |

| 2 | DeepSeek-V4.1-Flash-Q2_K.gguf | pending conversion and validation |

| 3 | DeepSeek-V4.1-Flash-Q3_K_M.gguf | pending conversion and validation |

| 4 | DeepSeek-V4.1-Flash-Q4_K_M.gguf | pending conversion and validation |

| 5 | DeepSeek-V4.1-Flash-Q5_K_M.gguf | pending conversion and validation |

About DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash

is a multimodal sparse Mixture-of-Experts model. The upstream card describes a

763B-parameter model with a 552B backbone, a 1M-token maximum context, and

approximately 8B active parameters during prefill and 16B during decode. It

uses a 40-layer causal encoder-decoder design, Engram n-gram lookup tables,

hyper-connections, compressed sparse attention, and a vision encoder. See the

official model card

for the original architecture, reference implementation, and usage limits.

!DeepSeek-V4.1-Flash upstream agentic performance benchmark

*Upstream DeepSeek-V4.1-Flash agentic performance benchmark; this image and

its scores are from the official model card,

not from these GGUF files.*

The first GGUF package targets text inference. Vision, multimodal projector,

and speculative MTP components will not be claimed unless they are separately

converted and validated with the actual runtime.

This is a quantization-only release. No training, fine-tuning, merging, or

model capability modification other than GGUF conversion and quantization is

performed here.

Fidelity measurements

This section will be added after a fixed hold-out comparison against a valid

reference has completed. No fidelity percentages are claimed yet.

Quick start

Commands will be added here with the exact published filename after the first

GGUF passes load and generate smoke tests.

Reproducibility and validation

The upstream checkpoint is deepseek-ai/DeepSeek-V4.1-Flash at revision

dba1be0a40aa45a94ad051997016db3960a90277. Its index reports 48 safetensors

shards and a total source size of 510,286,023,000 bytes. The checkpoint uses

the official mixed FP8/FP4 representation declared by its quantization_config;

it is not a BF16 GGUF input and no community GGUF is used as a quantization

source.

Conversion and runtime support use the DeepSeek-V4.1 work in

vcruz305/llama.cpp, commit

5210c7c5ed61dddaee6ed476623abf4b63093d16, with the V4.1 converter/runtime

path documented in the local reproducibility manifest. Each published file

will include a SHA-256 checksum and will be listed only after a real load and

generation smoke test.

Raw conversion, imatrix, quantization, smoke-test, fidelity, and benchmark

logs remain local under reports/ and are not uploaded to the Hub.

License and attribution

The upstream model is released under the MIT License. Preserve upstream

attribution and the included LICENSE when redistributing these

derivative artifacts.

These are community GGUF quantizations, not an official DeepSeek release or

endorsement.