Cipher-Abliterated

Cipher-Abliterated is an adaptive conversational AI model published on Ollama at vatistasdim/Cipher-Abliterated. It was made and trained by Dimitris Vatistas with https://dvatistas.vercel.app/. It is designed for flexible interaction, creative reasoning, structured output, and technical workflows through local Ollama deployment.

Ollama model page: https://ollama.com/vatistasdim/Cipher-Abliterated

Quick Start

Install Ollama, then run:

ollama pull vatistasdim/Cipher-Abliterated
ollama run vatistasdim/Cipher-Abliterated

ollama run will download the model automatically if it is not already present.

Model Snapshot

Field	Value
Ollama namespace	vatistasdim/Cipher-Abliterated
Default tag	latest
Published size	2.0 GB
Input	Text
Catalog context window	128K
Recommended runtime context	2048 tokens
Recommended temperature	0.85
Last checked	2026-05-25

The Ollama catalog lists a 128K model context window. The model card recommends using a 2048-token runtime context for efficient local sessions.

Architecture Diagram

flowchart TD
    subgraph UserLayer["User / App Layer"]
        A["Prompt, chat history, brainstorm, or tool request"]
        B["CLI, desktop app, script, or REST client"]
    end

    subgraph RuntimeLayer["Local Ollama Runtime"]
        C["Model selector: vatistasdim/Cipher-Abliterated:latest"]
        D["Chat template and stop-token handling"]
        E["Runtime context: recommended num_ctx 2048"]
        F["Streaming or non-streaming response mode"]
    end

    subgraph ModelLayer["Cipher-Abliterated Model Profile"]
        G["Identity: Cipher-Abliterated"]
        H["Architecture: llama"]
        I["Parameters: 3.2B"]
        J["Embedding length: 3072"]
        K["Native context window: 131072 tokens"]
        L["Quantization: Q4_K_M"]
    end

    subgraph ControlLayer["Flexible Generation Controls"]
        M["temperature: 0.85"]
        N["Output target: adaptive, broad, exploratory"]
        O["Good for options, rewrites, brainstorming, comparisons"]
    end

    subgraph OutputLayer["Response Layer"]
        P["Markdown, lists, comparison tables, creative drafts, plans"]
        Q["Flexible local answer"]
    end

    A --> B --> C --> D --> E --> F
    F --> G
    F --> H
    F --> I
    F --> J
    F --> K
    F --> L
    G --> M
    H --> M
    I --> M
    J --> M
    K --> M
    L --> M
    M --> N --> O --> P --> Q

Plain-text architecture map:

User prompt / chat history
        |
        v
Client layer
CLI, app, script, REST request
        |
        v
Local Ollama runtime
model: vatistasdim/Cipher-Abliterated:latest
context: recommended 2048 tokens
streaming: supported
        |
        v
Cipher-Abliterated profile
identity: Cipher-Abliterated
architecture: llama
parameters: 3.2B
embedding length: 3072
native context: 131072 tokens
quantization: Q4_K_M
        |
        v
Flexible generation controls
temperature: 0.85
target: adaptive, broad, exploratory
        |
        v
Output
Markdown, comparison tables, rewrites, options, drafts, plans

Architecture Details

Cipher-Abliterated is configured as a flexible local text model for broader conversation, ideation, and exploratory technical work. Its runtime profile is less restrictive than Cipher, which helps it produce more varied answers and multiple solution paths.

Layer	Role
Ollama runtime	Handles local model loading, chat requests, streaming, and API access.
Cipher-Abliterated model profile	Applies model identity, sampling behavior, stop tokens, and runtime context.
Context window	Supports a large catalog context window, with 2048 tokens recommended for fast daily use.
Quantized weights	Q4_K_M quantization keeps the model practical for consumer hardware while preserving useful reasoning quality.
Sampling profile	temperature 0.85 gives broader variation for brainstorming and creative exploration.
Tool-ready output	Supports completion and tool-capable workflows through Ollama-compatible clients.

Full Technical Profile

Component	Cipher-Abliterated Detail
Model name	Cipher-Abliterated
Ollama tag	vatistasdim/Cipher-Abliterated:latest
Creator	Dimitris Vatistas
Website	https://dvatistas.vercel.app/
Published size	2.0 GB
Architecture family	llama
Parameter scale	3.2B
Quantization	Q4_K_M
Input mode	Text
Output mode	Completion and tool-capable text
Native context window	131072 tokens
Recommended daily context	2048 tokens
Embedding length	3072
Temperature	0.85
Primary behavior	Adaptive, flexible, exploratory, creative
Best output formats	Brainstorm lists, comparison tables, rewrites, outlines, options, draft plans

Behavior Profile

Cipher-Abliterated is the more open model in the Cipher pair. It is intended to explore more possibilities, produce broader answers, and adapt its tone and structure to the prompt. The higher temperature makes it the better choice for:

• brainstorming
• rewrite options
• creative drafts
• comparing multiple approaches
• research outlines
• broad technical planning
• exploratory local-agent workflows

Benchmark Profile

Benchmark results depend on hardware, prompt size, context length, and Ollama settings. Cipher-Abliterated is tuned for flexible output while staying small enough for practical local use.

Area	Cipher-Abliterated Profile	What This Means
Local speed	High for a 3.2B-class model	Good for interactive chat, brainstorming, and local app workflows.
Memory use	Low to moderate	Designed to run on consumer machines without a large GPU requirement.
Answer precision	Moderate to high	Can answer technical prompts, but is intentionally less narrow than Cipher.
Creativity	High	Better for ideation, rewriting, alternatives, and exploratory reasoning.
Long-context work	Strong when context is increased	Start at 2048 tokens, then raise context for larger documents or logs.
Structured output	Strong with clear prompting	Ask for tables, bullets, JSON-shaped output, or explicit sections.

Local Benchmark Snapshot

These are single local smoke-test numbers from the same machine and a short prompt. They are useful for relative runtime feel, not as universal benchmark claims. No quality score is implied by token speed.

Benchmark prompt: Write exactly six concise bullets comparing local AI assistants for coding, summarization, and brainstorming.

Benchmark options: num_ctx 2048, num_predict 140, temperature 0.2.

Model	Installed size	Eval tokens	Total time	Generation speed
vatistasdim/Cipher-Abliterated:latest	2.0 GB	140	4.32 s	38.46 tok/s
vatistasdim/Cipher:latest	2.0 GB	137	12.60 s	32.19 tok/s
hf.co/bartowski/Qwen2.5-3B-Instruct-GGUF:Q4_K_M	1.9 GB	78	10.31 s	34.79 tok/s
phi3:mini	2.2 GB	140	8.27 s	33.17 tok/s
gemma:2b	1.7 GB	140	5.55 s	46.54 tok/s
dolphin-phi:latest	1.6 GB	140	8.81 s	38.00 tok/s
huihui_ai/falcon3-abliterated:3b	2.0 GB	140	12.53 s	36.29 tok/s

Near-2GB Model Comparison

Model	Size Class	Main Feel	Cipher-Abliterated Difference
gemma:2b	1.7 GB local model	Fast, lightweight general chat	Cipher-Abliterated is broader and more exploratory for ideation and planning.
phi3:mini	2.2 GB local model	Compact reasoning and instruction following	Cipher-Abliterated is looser and better for generating multiple options.
dolphin-phi:latest	1.6 GB local model	Lightweight conversational assistant	Cipher-Abliterated is more focused on adaptive brainstorming and technical exploration.
hf.co/bartowski/Qwen2.5-3B-Instruct-GGUF:Q4_K_M	1.9 GB local model	General instruction model with broad use	Cipher-Abliterated has a more explicit creative and exploratory assistant profile.
huihui_ai/falcon3-abliterated:3b	2.0 GB local model	Flexible 3B-class generation	Cipher-Abliterated is positioned for local assistant workflows and structured ideation.
vatistasdim/Cipher:latest	2.0 GB Cipher variant	Precise, concise, stricter	Cipher-Abliterated is the better option when variety and wider exploration matter.

Request Flow

sequenceDiagram
    participant User
    participant Client as "CLI, app, or script"
    participant Ollama as "Local Ollama runtime"
    participant Model as "Cipher-Abliterated model"

    User->>Client: "Send prompt"
    Client->>Ollama: "Chat request with model vatistasdim/Cipher-Abliterated"
    Ollama->>Model: "Apply model settings and context"
    Model-->>Ollama: "Generated response"
    Ollama-->>Client: "Response payload or streamed tokens"
    Client-->>User: "Flexible, structured answer"

Strengths

• More adaptive conversational behavior.
• Strong reasoning and problem-solving ability.
• Useful for creative ideation, research, coding, and exploration tasks.
• Structured output support when the prompt asks for a clear format.
• Compatible with local agent and automation systems through Ollama.

Local API Usage

Start the Ollama service, then call the chat API:

curl http://localhost:11434/api/chat \
  -d '{
    "model": "vatistasdim/Cipher-Abliterated",
    "messages": [
      { "role": "user", "content": "Brainstorm three approaches for a local coding assistant." }
    ],
    "stream": false,
    "options": {
      "temperature": 0.85,
      "num_ctx": 2048
    }
  }'

Python:

from ollama import chat

response = chat(
    model="vatistasdim/Cipher-Abliterated",
    messages=[{"role": "user", "content": "Hello!"}],
)

print(response.message.content)

JavaScript:

import ollama from "ollama";

const response = await ollama.chat({
  model: "vatistasdim/Cipher-Abliterated",
  messages: [{ role: "user", content: "Hello!" }],
});

console.log(response.message.content);

Application Launch Examples

ollama launch claude --model vatistasdim/Cipher-Abliterated
ollama launch codex-app --model vatistasdim/Cipher-Abliterated
ollama launch openclaw --model vatistasdim/Cipher-Abliterated
ollama launch codex --model vatistasdim/Cipher-Abliterated
ollama launch opencode --model vatistasdim/Cipher-Abliterated

Best Fit

Use Cipher-Abliterated when you want a flexible local assistant for:

• Brainstorming
• Creative writing support
• Comparing options
• Exploratory technical planning
• Rewriting and expanding ideas
• Research-style outlines
• Agent workflows where variety is useful

For stricter formatting and more precise answers, use Cipher. For open-ended thinking and broader response variation, use Cipher-Abliterated.