Instructions to use Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M

Use Docker

docker model run hf.co/Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M

Ollama
How to use Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF with Ollama:
```
ollama run hf.co/Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M
```
Unsloth Desktop

How to use Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF with Docker Model Runner:
```
docker model run hf.co/Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M
```

Lemonade

How to use Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Solstice-AI Banner

DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF

Official Solstice-AI Flagship Release • 552B MoE (384 Experts) • Protium-1.0 Objective-Gated Cognitive Architecture • 1M Context • Multimodal Vision

Base Model by DeepSeek AI • Permanent Weight-Level Abliteration by DealignAI • GGUF Quantization, Multimodal Integration & Cognitive Architecture by Solstice-AI

Executive Summary

Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF is the premier, unconstrained GGUF release of DeepSeek's frontier DeepSeek-V4.1-Flash causal decoder architecture. This release integrates permanent weight-level abliteration, full multimodal vision support via DeepSeek-V4.1-Flash-mmproj-BF16.gguf, a 1,048,576-token (1M) context window, and Solstice-AI's Protium-1.0 Objective-Gated Cognitive Architecture.

Unlike stock conversational models burdened by synthetic guardrail refusals, cognitive stalling, and unconstrained token loops, this release unlocks raw frontier intelligence for advanced security research, deep system penetration analysis, adversarial logic verification, unrestricted creative writing, and autonomous agent systems.

Key Capabilities & Architectural Highlights

Permanent Weight-Level Abliteration (Zero Refusals): Refusal direction vectors have been surgically abliterated at the weight level across transformer blocks. The model delivers 0% refusal across benchmark stress-suites without moralizing, preachy disclaimers, or conversational hedging, while preserving 100% of MMLU, mathematical reasoning, agentic coding, and DSpark speculative decoding capabilities.
Extreme Frontier MoE Architecture: Features a 552B total parameter backbone with 384 routed experts (top-6 active) + 1 shared expert (8B to 16B active per token), 4-channel Hyper-Connections, CSA2 sparse attention, and Engram n-gram memory lookup tables.
Multimodal Vision Integration: Ships with DeepSeek-V4.1-Flash-mmproj-BF16.gguf (0.91 GB), enabling high-resolution visual processing, technical diagram architecture auditing, OCR, and GUI agent workflows.
Protium-1.0 Objective-Gated Cognitive Architecture: Integrated directly into chat_template.jinja at the repo root. Powered by the [L10] Proteus Dynamic Adaptive Engine (default) with zero forced reasoning quotas, forward-only target anchoring, and specialized multi-agent convergence modes ([swarm] for 20-perspective hypothesis exploration and [council] for adversarial red-team panels).
1M-Token Context Window (YaRN): Native support for up to 1,048,576 tokens via YaRN RoPE factor 16, allowing entire codebases, complete documentation archives, and days of multi-agent interactions to reside in memory simultaneously.

Shard Topology & Memory Budgets

Due to the 552B parameter scale, checkpoints are split into clean GGUF shards compatible with standard llama.cpp shard loaders:

Quant Tier	Shards	Total Size	Practical Memory Budget (RAM / VRAM)	Operational Profile
`Q4_K_M` Suite	3 Shards	~396.83 GB	~450 GB+	⭐️ Recommended Production Baseline — Optimal balance of reasoning retention and inference throughput. Ideal for multi-GPU servers (e.g. 8×A100/H100 80GB, 4×H200 141GB, or dual Grace Hopper / GB200 systems).
`Q8_0` Suite	4 Shards	~473.07 GB	~540 GB+	Reference Grade Precision — Zero discernible quantization loss compared to original weights. Ideal for frontier research clusters and enterprise multi-node deployments.
`mmproj-BF16`	1 File	0.91 GB	Included in host	DeepSeek-ViT Vision Projector — Required for multimodal image comprehension.
`chat_template.jinja`	1 File	33.9 KB	—	Protium-1.0 Native Template — Full Jinja template implementing Objective-Gated cognition.

Shard Manifest

Q4_K_M Files:
- DeepSeek-V4.1-Flash-UNCENSORED-Q4_K_M-00001-of-00003.gguf (156.11 GB)
- DeepSeek-V4.1-Flash-UNCENSORED-Q4_K_M-00002-of-00003.gguf (165.56 GB)
- DeepSeek-V4.1-Flash-UNCENSORED-Q4_K_M-00003-of-00003.gguf (75.16 GB)
Q8_0 Files:
- DeepSeek-V4.1-Flash-UNCENSORED-Q8_0-00001-of-00004.gguf (165.85 GB)
- DeepSeek-V4.1-Flash-UNCENSORED-Q8_0-00002-of-00004.gguf (103.10 GB)
- DeepSeek-V4.1-Flash-UNCENSORED-Q8_0-00003-of-00004.gguf (97.28 GB)
- DeepSeek-V4.1-Flash-UNCENSORED-Q8_0-00004-of-00004.gguf (106.84 GB)
Vision Projector:
- DeepSeek-V4.1-Flash-mmproj-BF16.gguf (0.91 GB)

Loading Note: In llama.cpp, you only need to pass the first shard (...-00001-of-...gguf); the runtime automatically locates and stitches the remaining shards in the sequence.

Protium-1.0 Objective-Gated Cognitive Architecture

Built natively into chat_template.jinja. Full protocol documentation: Solstice-AI/Protium-1.0.

Core Mechanics ($^1\text{H}$ Zero-Waste Protocol)

Token 1 Target Anchoring: At token 1, the model isolates and locks the exact termination criteria — defining what a completed, verified solution requires before executing.
Forward-Only Deduction: Deductive steps progress strictly forward. Reflexive rhetorical pauses ("let's see", "now let me check") are systematically suppressed.
Anti-Pivot Invariant: The model is forbidden from emitting hesitation pivots ("Wait...", "On second thought...") unless a concrete mathematical, programmatic, or factual contradiction is explicitly isolated.
Objective-Gated Early Exit: The moment verification criteria are confirmed, deliberation terminates immediately and the final response is delivered.

Cognitive Spectrum & Multi-Agent Swarm Modes

The engine defaults to [L10] Proteus Dynamic Adaptive, auto-tuning reasoning impedance to the workload. Prefix any prompt with an invocation marker:

Level	Identifier	Archetype	Operational Focus	Workload Profile
`[L10]`	`proteus` / `dynamic`	Dynamic Adaptive (Default)	Fluid impedance matching	Auto-calibrated depth; collapses on lookup/boilerplate, expands on proofs and complex architectures.
`[L0]`	`mortal` / `regular`	Instant Execution	Thinking suppressed	Pure conversational dialogue, zero-overhead classification, streaming API endpoints.
`[L1]`	`hermes`	Rapid Instinct	High-velocity verification	Single-pass sanity checks, syntax validation, arithmetic confirmation.
`[L2]`	`apollo`	Disciplined Deduction	Linear constraint trace	Parameter mapping, rule following, structured schema validation.
`[L3]`	`artemis`	Boundary & Failure Audit	Edge vulnerability probe	Null/empty states, numeric limits, concurrency race condition auditing.
`[L4]`	`athena`	Strategic Architecture	Modular synthesis	Multi-step refactoring, clean API interface design, trade-off optimization.
`[L5]`	`prometheus`	Defensive Engineering	Second-order impact modeling	Fault tolerance, graceful degradation, scaling lifecycle resilience.
`[L6]`	`solstice`	Deep Systemic Derivation	Multi-hypothesis synthesis	Root cause isolation in legacy codebases, algorithmic breakthroughs.
`[L7]`	`epimetheus`	Adversarial Stress-Test	Active disproof search	Aggressive edge-case fuzzing, red-teaming, boundary testing.
`[L8]`	`hyperion`	Exhaustive Combinatorial	Full solution-space traversal	Comprehensive state-machine enumeration, non-trivial mathematical optimization.
`[L9]`	`oracle`	Proof-Grade Authority	Formal verification	Exhaustive axiomatic proofs, mission-critical zero-placeholder implementations.
`[swarm]`	`swarm`	20-Perspective Swarm	Divergent synthesis	Spawns an internal multi-agent swarm analyzing the problem across 20 distinct domain perspectives before synthesizing a verified consensus.
`[council]`	`council`	Adversarial Red-Team Panel	Dialectical debate	Simulates an adversarial council of specialized experts debating trade-offs, attack vectors, and edge failures until absolute convergence.

Example Invocation:

> [swarm] Analyze this zero-day memory corruption vulnerability in the kernel network stack and derive both an exploit proof-of-concept and an unbypassable patch.

Quickstart & Serving Guide

1. Download Checkpoint via `hf` CLI

Download the Q4_K_M shards, vision projector, and native chat template:

hf download Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF \
  --include "DeepSeek-V4.1-Flash-UNCENSORED-Q4_K_M-*" \
            "DeepSeek-V4.1-Flash-mmproj-BF16.gguf" \
            "chat_template.jinja" \
  --local-dir ./dsv41-uncensored-protium

2. Run with `llama.cpp` (`llama-server`)

Note on Runtime: DeepSeek-V4.1 utilizes the deepseek41 architecture (upstream PR ggml-org/llama.cpp#28696). Ensure your llama.cpp build supports the DeepSeek-V4.1 loader, Engram memory tables, and CSA2 sparse attention.

llama-server \
  --model ./dsv41-uncensored-protium/DeepSeek-V4.1-Flash-UNCENSORED-Q4_K_M-00001-of-00003.gguf \
  --mmproj ./dsv41-uncensored-protium/DeepSeek-V4.1-Flash-mmproj-BF16.gguf \
  --chat-template-file ./dsv41-uncensored-protium/chat_template.jinja \
  --host 0.0.0.0 --port 8080 \
  --ctx-size 131072 \
  -ngl 99 \
  --temp 0.6 \
  --top-p 0.95

3. Run with Ollama

Create a Modelfile:

FROM ./DeepSeek-V4.1-Flash-UNCENSORED-Q4_K_M-00001-of-00003.gguf

PARAMETER temperature 0.6
PARAMETER top_p 0.95
PARAMETER num_ctx 65536
PARAMETER stop "<|endoftext|>"
PARAMETER stop "<|im_end|>"

TEMPLATE """{{ if .System }}<｜begin of sentence｜><｜system｜>
{{ .System }}{{ end }}{{ if .Prompt }}<｜User｜>
{{ .Prompt }}{{ end }}<｜Assistant｜><think>
{{ .Response }}<｜end of sentence｜>"""

Build and run:

ollama create dsv41-uncensored-protium -f Modelfile
ollama run dsv41-uncensored-protium

4. Recommended Sampling Parameters

Scenario	Temperature	Top-p	Top-k	Max Output Tokens
Complex Reasoning & Coding	0.6	0.95	40	131,072
Offensive Security & Red-Teaming	0.7	0.95	50	131,072
Uncensored Creative & Brainstorming	0.9	0.98	-	65,536
Fast Classification & Direct Answers	0.2	0.85	20	8,192

Architectural Specifications

Property	Value
Model Name	DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF
Base Architecture	DeepSeek-V4.1 (`DeepseekV41ForCausalLM`)
Base Weights	`Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-FP8` / `deepseek-ai/DeepSeek-V4.1-Flash`
Total Backbone Parameters	552 Billion
Active Parameters	8 Billion (standard) to 16 Billion (extended) per token
Layers	40 Transformer blocks
Hidden Dimension	5,120
Attention Heads	64 query heads, 1 key/value head (Multi-Head Latent Attention - MLA)
MoE Experts	384 routed experts (top-6 active) + 1 shared expert
Intermediate MoE Size	2,304
Residual Topology	4-channel Hyper-Connections
Sparse Attention	CSA2 with Engram n-gram memory lookups
Speculative Decoding	DSpark MTP draft head compatible
Native Context Length	1,048,576 tokens (1M) via YaRN RoPE factor 16
Vocabulary Size	129,280
Multimodal Vision	DeepSeek-ViT with 2D-RoPE + pixel unshuffle (`mmproj` included)
Alignment Status	Permanent Weight-Level Abliterated (Dealigned / Uncensored)

Responsible Use & Security Disclaimer

This checkpoint is permanently de-aligned and abliterated. All safety guardrails, refusal filters, and content moderation boundaries have been surgically removed at the weight level.

Intended Purpose: Designed strictly for authorized cybersecurity research, red-teaming, vulnerability assessment, defense engineering, AI interpretability experiments, and unconstrained creative writing.
Operator Liability: The user assumes full operational, legal, and ethical responsibility for all inputs submitted to and outputs generated by this model. Solstice-AI and affiliated contributors assume no liability for misuse.

Citations & Acknowledgments

DeepSeek AI for developing the landmark DeepSeek-V4.1-Flash architecture and open-weights foundation.
DealignAI for the weight-level abliteration and de-alignment research.
vcruz305 for the DeepSeek-V4.1 GGUF conversion PR (ggml-org/llama.cpp#28696).
Solstice-AI for the Protium-1.0 Objective-Gated Cognitive Architecture and curated UltraOptimised quantization suite.

@misc{deepseek2026v41flash,
  title        = {DeepSeek-V4.1-Flash: Pushing the Frontier of Efficient Multimodal MoE},
  author       = {DeepSeek AI},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash}}
}

@misc{solstice2026protium,
  title        = {DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF},
  author       = {Solstice-AI},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF}}
}

Crafted with precision by Solstice-AI
Objective-Gated Cognition • Zero-Waste Reasoning • Ultra-Optimised Compute

Downloads last month: 18,961

GGUF

Model size

748B params

Architecture

deepseek41

Hardware compatibility

4-bit

8-bit

DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF

Official Solstice-AI Flagship Release • 552B MoE (384 Experts) • Protium-1.0 Objective-Gated Cognitive Architecture • 1M Context • Multimodal Vision

Executive Summary

Key Capabilities & Architectural Highlights

Shard Topology & Memory Budgets

Shard Manifest

Protium-1.0 Objective-Gated Cognitive Architecture

Core Mechanics ($^1\text{H}$ Zero-Waste Protocol)

Cognitive Spectrum & Multi-Agent Swarm Modes

Example Invocation:

Quickstart & Serving Guide

1. Download Checkpoint via hf CLI

2. Run with llama.cpp (llama-server)

3. Run with Ollama

4. Recommended Sampling Parameters

Architectural Specifications

Responsible Use & Security Disclaimer

Citations & Acknowledgments

1. Download Checkpoint via `hf` CLI

2. Run with `llama.cpp` (`llama-server`)