Downloads · 30 days
31
6% of all-time downloads
hampsonw/DeepSeek-V4-Flash-0731-DIQ3S-GIQ1S-UIQ1M-GGUF
DeepSeek-V4-Flash-0731-DIQ3S-GIQ1S-UIQ1M-GGUF is a text generation model from hampsonw. Use it when you need the model to write or continue text. It is set up for llama.cpp. The card lists the license as mit.
This is an experimental role-spliced GGUF for deepseek-ai/DeepSeek-V4-Flash-0731. It is not an official DeepSeek release.
Downloads · 30 days
31
6% of all-time downloads
All-time downloads
500
Public
Repo size
86.7 GB
Likes
1
Public
Click a slice to open those files.
.gguf86.7 GB · 100%
From the Hugging Face model README
This is an experimental role-spliced GGUF for deepseek-ai/DeepSeek-V4-Flash-0731. It is not an official DeepSeek release.
The file combines already-quantized tensor payloads from four revision-pinned GGUFs. No tensor was dequantized or requantized during assembly.
| Routed-expert role | Tensors | Type | Donor |
|---|---|---|---|
ffn_down_exps.weight | 43 | IQ3_S | TacoTakumi |
ffn_gate_exps.weight | 43 | IQ1_S | 6block |
ffn_up_exps.weight | 43 | IQ1_M | AtomicChat |
The other 1,199 tensors come from antirez/deepseek-v4-gguf, including its Q8/F16/F32 attention, shared-expert, router, output, and auxiliary tensors.
Complete tensor-type census:
| Type | Tensors |
|---|---|
F32 | 492 |
F16 | 359 |
Q8_0 | 345 |
IQ3_S | 43 |
IQ1_S | 43 |
IQ1_M | 43 |
I32 | 3 |
DeepSeek-V4-Flash-0731-DIQ3S-GIQ1S-UIQ1M.gguf7cd6c76a910df9a05567e07ad3955e93ec9373fddc2c5732382d9214261593e9c0c5c28c4cc48e3845368686c63649e462afb9a07ce597cf4fc306dbb9495de1splice-manifest.json records every source repository revision, LFS SHA-256, source byte range, tensor type, shape, and output offset. verification.json records structural verification and the runtime smoke test.
The assembled file passed two checks:
llama-cli from llama.cpp revision 2f56fc3431f47fe042bf3825e4d5523bdddda993 loaded the model, initialized a 128-token context, evaluated Hello, generated one token, and exited successfully.This splice has not been quality-benchmarked. The three expert donors used different importance matrices and calibration corpora. GGUF tensor blocks are independently decodable, but combining them does not prove that the resulting model preserves the quality of any donor release.
Treat this as an experimental candidate until it has been compared with the Antirez, Unsloth, and uniform baselines on the same KL/perplexity and task harness.
The file does not include an MTP or DSpark sidecar.
Use a llama.cpp build with deepseek4 support. The smoke test used revision 2f56fc3.
llama-cli \
-m DeepSeek-V4-Flash-0731-DIQ3S-GIQ1S-UIQ1M.gguf \
-c 4096 \
-p "Hello"
Configure GPU offload and context size for your hardware. The GGUF alone occupies 80.76 GiB before runtime buffers and KV cache.
The base model and inherited license come from deepseek-ai/DeepSeek-V4-Flash-0731. Quantized payloads are credited to Antirez, TacoTakumi, 6block, and AtomicChat at the pinned revisions above.
License: MIT.