---

license: other license_name: mit-weights-and-runtime-with-nvidia-cuda-redistributables license_link: 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 / Windows

Renderoverse's portable packaging of an existing local AtomicChat IQ2_XXS setup, with the separate Engram table, MTP/vision sidecar, BF16 vision tower and qualified native Windows runtime. This is a runtime integration/repack, not a new training run or a claim of better benchmark quality than upstream. No new censorship modifications were made.

Distribution: Renderoverse release page. The complete torrent contains all four weights, runtime.zip, notices and checksums. Hugging Face may contain only a subset: check its actual file list before downloading. This repository is held private for owner review; its visibility must not be changed without owner approval.

What you download

File, under 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

Weights total 255,467,883,840 bytes = 255.47 GB = 237.92 GiB. runtime.zip adds about 0.49 GB. Python and CUDA PyTorch are installed separately from their official package indexes; those downloads need several more GB. The runtime archive intentionally includes an EXE, DLLs and Python/PowerShell scripts. Inspect it before executing; it is not merely model data.

Requirements

• Windows x64, CPython 3.12 x64, and an NVIDIA driver compatible with CUDA 13.
• This exact native build comes from an RTX 5090 (32 GB) setup. Other GPUs/CPUs have not been qualified for this package; this is not a generic llama.cpp/Ollama import.
• At least 11 logical CPUs available for the fixed six-worker affinity mask (0x555). The launcher fails closed if its CPU job limits cannot be applied.
• 92 GiB free system RAM at startup (60 GiB arena plus 32 GiB headroom); 128 GiB+ installed RAM is recommended. Do not use the pagefile as a substitute.
• 29,000 MiB free on GPU 0 at startup. Do not run alongside training/rendering or another loaded model.
• Fast SSD; reserve at least 270 GiB free for release, environment and working space. Retaining additional downloads/cache needs extra space.

The Engram table streams from SSD; startup, context length and speed depend strongly on RAM/SSD/GPU resources. Configured context is 262,144, not a guaranteed tested maximum for every machine. Sustained 60–80 tok/s and quality superiority have not been established.

Install and run

1. Download the entire torrent. Keep all four GGUFs in weights/.
2. Extract runtime.zip into this same release directory, not into another nested folder. You should have launch.py, install.ps1, start.ps1, runtime/ and weights/ side by side.
3. Install Python 3.12 x64 and your supported NVIDIA driver. On Windows installations that need it, install the official Microsoft Visual C++ x64 runtime.
4. Open PowerShell in the release folder and run the reviewed scripts:

.\install.ps1
.\start.ps1

Installation creates a folder-local .venv; it installs pinned base requirements and torch==2.10.0 from the official CUDA 13 index. If your PowerShell policy blocks scripts, follow your own organization's signing/unblocking policy; the package does not change execution policy or security settings.

For a no-model/no-GPU-load check:

.\start.ps1 --check
# Optional: reread and SHA-256-check every weight (255 GB of SSD reads):
.\start.ps1 --check --verify

The server listens on 127.0.0.1:18101 only. Startup prints the private per-install API-key file location, not the key. The key/config/cache/logs live under %LOCALAPPDATA%\Renderoverse\DeepSeek-Native-Vision\<installation-id>\, never in the downloadable release. Do not share that state directory. The built-in local web chat is at http://127.0.0.1:18101/; API base is http://127.0.0.1:18101/v1.

Stop with Ctrl+C in the server window. Closing the window also closes its Windows job and descendants. The package never closes other applications or restarts Harness. It preserves the six-worker affinity/BelowNormal job, BLAS=1, a CPU startup check and emergency CPU ceiling; an additional RAM guard stops only this release tree after two low-memory samples. A guard exit is an intentional stop, not an instruction to weaken the guard. CPU telemetry/receipts are local to this installation.

OpenAI-compatible examples

Use the key file path printed by startup. These snippets use the installed requests package; run them with .venv\Scripts\python.exe. They are examples, not benchmark results.

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"])

For an image, replace body["messages"] before sending:

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}}
]}]

Function tools use the normal OpenAI tools schema. The server translates DeepSeek's DSML tool-call format into message.tool_calls; your client decides whether/how to execute tools. Reasoning effort supports none, low, medium, high, max and numeric 1–100. Large images are limited by the configured 2048-token vision budget.

Verification status and limitations

The release was tested from a relocated extracted archive in an independent Python environment: runtime checksums/imports, CPU job readback, Unicode tokenizer round trips, text/image prompt formatting, lazy vision-helper paths, vision routing-bias compatibility, cache binding, cuBLAS DLL loading, and authenticated mock HTTP text/tool APIs passed. The mock engine generates scripted responses; these checks do not prove inference quality. No new GPU inference/vision-output test was performed on this portable release because another training job was active. Treat this as an experimental release candidate, not a universally validated installer.

SHA-256 checksums are in SHA256SUMS.txt and PROVENANCE.json; the runtime also checks its own packaged file hashes on startup. --verify additionally checks the large weights. There are no original conversations, API keys, Harness configuration, .venv, or private disk caches in this release. Local prompt/cache state created after launch can contain private content.

Credits and licenses

Model: DeepSeek, uncensored weights: AtomicChat, GSQ-RCO/Engram: ISTA-DASLab, native engine/shims: Indra's Mirror Xolotl and Strata. Original rights/notices are retained. See NOTICE.md, PROVENANCE.json and runtime/licenses/ after extraction. NVIDIA DLLs are separately licensed under the included CUDA EULA; the whole archive is not exclusively MIT.

Torrent availability

Renderoverse stores the small torrent metadata, not these 255 GB of payload. Payload comes from seed peers. Availability and speed depend on online seeds, their upload bandwidth and network reachability; a listing alone is not cloud storage or a permanent seed. Keep the torrent seeding after download if you want to help distribution.