license: other
license_name: mit-weights-and-runtime-with-nvidia-cuda-redistributables
license_link: https://huggingface.co/zuraxe/DeepSeek-V4.1-Flash-1017-Uncensored-Native-Vision-Windows/blob/main/NOTICE.md
base_model:
- AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF
- ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF
pipeline_tag: image-text-to-text
tags: - gguf
- deepseek
- uncensored
- vision
- windows
- strata
DeepSeek V4.1 Flash 1017 Uncensored — Native Vision for RTX 5090 (Windows)
This package brings AtomicChat's uncensored DeepSeek V4.1 Flash 1017 together with the files and Windows runtime needed for text chat and image input. You can use the local chat page or connect a client through the OpenAI-compatible API.
The model weights come from the original authors. What's provided here is the runtime, setup scripts and instructions for putting the matching files together.
Choose your download
This Hugging Face repository contains about 0.49 GB of software and documentation. No model weights are reuploaded here. Follow the links below to get them from their authors.
If you'd rather download everything in one go, the full Renderoverse torrent includes all four model files and the runtime. It's 255.96 GB before installing Python dependencies. With the torrent, you can skip the shard-merging steps and go straight to Install and run.
Get the weights from their authors
1. Download the runtime
Extract runtime.zip into a new folder. Download prepare_atomic.py, gguf_merge_shards.py and UPSTREAM_WEIGHTS.json from this repository and put them in the same folder.
2. Download and merge the main model
Open AtomicChat's IQ2_XXS files and download all four files named DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-0000N-of-00004.gguf. Put them in a folder called atomic-shards. This setup uses IQ2_XXS, not IQ1_M or IQ2_XS.
With Python 3.12 installed, open PowerShell in the runtime folder and run:
py -3.12 .\prepare_atomic.py --shards .\atomic-shards
This checks the four downloads, merges them into weights/atomic-iq2_xxs.gguf without requantizing, and checks the finished file. It keeps the original shards and won't overwrite an existing output. The CPU limits also apply during the merge.
Allow about 129 GiB of extra SSD space, plus 2 GiB of headroom, for the merged file. If you keep the shards as well as the full setup, plan for over 400 GiB free. Add --plan-only to check paths and sizes without starting the merge.
3. Download the three companion files
Put these in weights/, alongside the merged model:
- DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf — the Engram table, which is read from your SSD.
- mmproj-DeepSeek-V4.1-Flash-1017-BF16.gguf — the vision tower.
- V41-MTP-Vision-1017.Q2_K.gguf — the MTP/vision companion file.
These links point to the specific versions used by this package. UPSTREAM_WEIGHTS.json lists their sizes and SHA-256 hashes, including the expected hash of the merged main model.
What you'll need
This build was put together for Windows x64 and an RTX 5090 with 32 GB VRAM. Other hardware hasn't been tested with this package. It uses its own native runtime; it isn't an Ollama or standard llama.cpp import.
- Python 3.12 x64 and an NVIDIA driver that supports CUDA 13.
- 92 GiB of free system RAM at startup. A machine with at least 128 GiB installed is recommended. Pagefile space doesn't count toward this check.
- 29,000 MiB of free VRAM on GPU 0. Close other GPU workloads before starting it.
- At least 11 logical CPUs available. The launcher uses six workers with a fixed affinity mask of
0x555. - A fast SSD with at least 270 GiB free for the assembled package, Python environment and working space. Keep more space free if you're retaining download copies or merging shards.
The four model files total 255.47 GB (237.92 GiB):
File in weights/ |
Bytes |
|---|---|
atomic-iq2_xxs.gguf |
138,238,892,608 |
DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf |
110,595,290,816 |
V41-MTP-Vision-1017.Q2_K.gguf |
5,661,492,320 |
mmproj-DeepSeek-V4.1-Flash-1017-BF16.gguf |
972,208,096 |
The runtime adds about 0.49 GB. Python dependencies need several more GB. Startup and generation speed depend on your SSD, RAM, GPU and context length. The configured context limit is 262,144 tokens; that full length hasn't been tested across different machines.
Install and run
Once you have the four model files, keep them in weights/ and extract runtime.zip into the same parent folder. The layout should look like this:
launch.py
install.ps1
start.ps1
runtime/
weights/
Install Python 3.12 x64 and a compatible NVIDIA driver. If Windows reports missing Visual C++ libraries, install the Microsoft Visual C++ x64 runtime.
Open PowerShell in that folder:
.\install.ps1
.\start.ps1
The installer creates a local .venv and installs the pinned dependencies, including torch==2.10.0 from the official CUDA 13 package index. The download includes executable files and scripts, so read the scripts before running them. If PowerShell blocks them, use your normal script-signing or unblocking process; the installer doesn't change your execution policy.
To check the setup without loading the model or using the GPU:
.\start.ps1 --check
To also verify every model file, run start.ps1 --check --verify. That reads all 255 GB of weights, so it takes longer.
When the server is running, open the local chat page. The API address is http://127.0.0.1:18101/v1. Both are local to your computer.
Startup tells you where to find your API key. The key, configuration, logs and cache are stored under %LOCALAPPDATA%\Renderoverse\DeepSeek-Native-Vision\<installation-id>\. Keep that folder to yourself; it may contain chat content.
Press Ctrl+C in the server window to stop it. The launcher limits CPU use and checks available memory. If a resource check stops the server, free up resources before trying again. It won't start without its CPU limits in place.
Connect your own client
Use the API-key path printed at startup. Run this example with .venv\Scripts\python.exe:
import pathlib, requests
key = pathlib.Path(r"PATH_PRINTED_AT_STARTUP\api-key.txt").read_text().strip()
headers = {"Authorization": "Bearer " + key}
model = requests.get("http://127.0.0.1:18101/v1/models", headers=headers).json()["data"][0]["id"]
body = {"model": model, "messages": [{"role": "user", "content": "Explain a rainbow."}],
"max_tokens": 256, "reasoning_effort": "none"}
r = requests.post("http://127.0.0.1:18101/v1/chat/completions", headers=headers, json=body, timeout=900)
r.raise_for_status()
print(r.json()["choices"][0]["message"])
To send an image, replace body["messages"] with this before making the request:
import base64
image = base64.b64encode(pathlib.Path("example.png").read_bytes()).decode()
body["messages"] = [{"role": "user", "content": [
{"type": "text", "text": "Describe what is visible in this image."},
{"type": "image_url", "image_url": {"url": "data:image/png;base64," + image}}
]}]
Tool calls use the OpenAI tools format and are returned in message.tool_calls; your client handles running the tools. Reasoning effort accepts none, low, medium, high, max, or a number from 1 to 100. The vision token budget is set to 2048.
Testing
The packaging checks passed, including file hashes, Python imports, CPU limits, tokenizer round trips, image-prompt formatting and API tests with a mock engine. Those API tests use scripted replies. Actual text and image generation haven't been re-tested end to end from this portable archive yet, and this package has no speed or quality benchmark results.
SHA256SUMS.txt and PROVENANCE.json record the download hashes. The runtime checks its own files when it starts; --check --verify also checks the weights.
Credits and licenses
The model is from DeepSeek, with uncensored weights from AtomicChat, GSQ-RCO and Engram files from ISTA-DASLab, and runtime work from Indra's Mirror Xolotl and Strata. The MTP/vision file comes from IndrasMirror.
See NOTICE.md, PROVENANCE.json and runtime/licenses/ for the original notices and terms. The NVIDIA DLLs have their own CUDA license; the archive as a whole isn't covered by MIT alone.
About the torrent
The torrent downloads the files from other people who are seeding them, not from a 256 GB cloud-storage copy on Renderoverse. Speed and availability depend on those seeds. If you can, leave your torrent client running after the download finishes so others can get it too.