--- title: DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF canonical_url: "https://www.modelscope.ai/models/Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF" md_url: "https://www.modelscope.ai/models/Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF.md" repository: Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF last_updated: 2026-10-30 license: other pipeline_tag: image-text-to-text tasks: - image-text-to-text model_type: - deepseek_v41 architectures: - DeepseekV41ForCausalLM base_model: - Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-FP8 base_model_relation: quantized language: - en - zh downloads: 0 stars: 0 tags: - solstice-ai - deepseek - deepseek-v4.1 - v4.1-flash - uncensored - abliterated - dealigned - protium - protium-1.0 - cognitive-architecture - dynamic-reasoning - objective-gated - zero-waste - speculative-decoding - dspark - 1m-context - multimodal - vision - mmproj - gguf - llama.cpp - ollama - vllm - sglang - q8_0 - q4_k_m --- # DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF > DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF - An open-source model by Solstice-AI on ModelScope. DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF is a image-text-to-text model on ModelScope. licensed under other. derived from Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-FP8. - **Repository**: Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF - **License**: other - **Tasks**: image-text-to-text - **Base model**: Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-FP8 - **Tags**: solstice-ai, deepseek, deepseek-v4.1, v4.1-flash, uncensored, abliterated, dealigned, protium, protium-1.0, cognitive-architecture, dynamic-reasoning, objective-gated, zero-waste, speculative-decoding, dspark, 1m-context, multimodal, vision, mmproj, gguf, llama.cpp, ollama, vllm, sglang, q8_0, q4_k_m - **Downloads**: 0 - **Stars**: 0 - **Last updated**: 2026-10-30 Source: https://www.modelscope.ai/models/Solstice-AI/DeepSeek-V4.1-Flash-UNCENSORED-Protium-GGUF ---

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

Solstice-AI License Format Alignment Architecture Context Vision Protium

--- ## 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 1. **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. 2. **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. 3. **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. 4. **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). 5. **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`](https://huggingface.co/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: ```text > [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: ```bash 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. ```bash 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`: ```dockerfile 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｜> {{ .Response }}<｜end of sentence｜>""" ``` Build and run: ```bash 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`](https://huggingface.co/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. ```bibtex @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