> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/Kismetkanceled/geniehelper/llms.txt
> Use this file to discover all available pages before exploring further.

# Ollama Models

> Seven locally-hosted LLM models running on Ollama, each optimized for specific tasks in the Genie Helper pipeline

Genie Helper uses a multi-model stack running on Ollama (port 11434) to balance performance, cost, and task specialization. All models run locally on the VPS with no external API calls.

## Model Inventory

| Model                         | Size  | Role                | Typical Use Case                                          |
| ----------------------------- | ----- | ------------------- | --------------------------------------------------------- |
| `dolphin3:8b-llama3.1-q4_K_M` | 4.9GB | Orchestrator        | Tool planning, ACTION tag emission, multi-step reasoning  |
| `dolphin-mistral:7b`          | 4.1GB | Uncensored writer   | NSFW captions, adult content drafts, fan messages         |
| `qwen-2.5:latest`             | 4.8GB | Primary agent       | AnythingLLM agent, code generation, JSON output           |
| `phi-3.5:latest`              | 2.2GB | Fallback classifier | Lightweight taxonomy tagging when scout-fast-tag fails    |
| `llama3.2:3b`                 | 2.0GB | Summarizer          | Quick summaries, metadata extraction, lightweight tasks   |
| `scout-fast-tag:latest`       | 1.2GB | Taxonomy classifier | Fast 6-concept taxonomy tagging (custom SmolLM fine-tune) |
| `bge-m3:latest`               | 2.4GB | Embeddings          | Document embeddings for RAG, semantic search              |

**Total RAM pinned**: \~21GB (4-5GB active at any time)

***

## Model Roles

### Primary Agent: `qwen-2.5:latest`

**Used by**: AnythingLLM workspace (default)

**Strengths**:

* Best code generation in the 7B class
* Reliable JSON output for structured tasks
* Good instruction following
* Handles most general-purpose agent tasks

**Weaknesses**:

* Tool calling is unreliable (hence Action Runner bypass)
* First token latency: \~33s in agent mode on CPU-only VPS
* Will stall on CPU-only hardware if context > 8K tokens

**When to use**:

* Default for all AnythingLLM chat interactions
* Code generation (scripts, flows, API integrations)
* Structured data extraction
* Technical Q\&A

***

### Orchestrator: `dolphin3:8b-llama3.1-q4_K_M`

**Used by**: Action Runner flow planning (optional)

**Strengths**:

* Excellent at multi-step planning
* Reliable ACTION tag emission
* Uncensored but still coherent for reasoning tasks
* Better tool selection than Qwen-2.5

**Weaknesses**:

* Slower inference (\~45s first token on CPU)
* Larger memory footprint (4.9GB)
* Currently stalls on production VPS (CPU-only)

**When to use**:

* Complex workflows requiring multi-step reasoning
* When you need better tool selection than Qwen-2.5
* **Requires GPU VPS or wait for hardware upgrade**

***

### Content Writer: `dolphin-mistral:7b`

**Used by**: Caption generation, fan message drafting

**Strengths**:

* Fully uncensored (no refusals on NSFW content)
* Creative, natural-sounding prose
* Good at matching tone/style from examples
* Fast inference (\~2-5s first token)

**Weaknesses**:

* Less structured than Qwen (worse for JSON)
* Can be overly verbose
* Not suitable for technical tasks

**When to use**:

* Drafting OnlyFans/Fansly captions
* Fan engagement messages (DMs, comments)
* Content descriptions and marketing copy
* Any NSFW text generation

**API endpoint**: `POST /api/captions/generate` (uses this model)

***

### Taxonomy Classifier: `scout-fast-tag:latest`

**Used by**: `taxonomy-tag` ACTION flow

**Strengths**:

* Custom fine-tune on 3,208 taxonomy tags
* Fast inference (\~500ms per image)
* Optimized for 6-concept classification
* Low memory footprint (1.2GB)

**Weaknesses**:

* Single-purpose model (only taxonomy)
* Requires fallback to `phi-3.5` on failure
* Limited to visual content classification

**When to use**:

* Auto-tagging uploaded media
* Batch classification of scraped content
* Content audit (NSFW detection, theme analysis)

**Output format**:

```json theme={null}
{
  "concepts": {
    "aesthetic": ["glamour", "luxury"],
    "activity": ["posing"],
    "mood": ["confident"],
    "setting": ["indoor", "bedroom"],
    "attire": ["lingerie"],
    "bodyType": ["athletic"]
  }
}
```

***

### Fallback Classifier: `phi-3.5:latest`

**Used by**: Taxonomy tagging when `scout-fast-tag` fails

**Strengths**:

* General-purpose vision model
* More robust on edge cases
* Can handle text + image input

**Weaknesses**:

* Slower than scout-fast-tag (\~3-5s)
* Less accurate on Genie Helper's specific taxonomy
* May refuse NSFW content (has guardrails)

**When to use**:

* Backup for failed scout-fast-tag calls
* Mixed content (text + image analysis)
* When you need explanations for classifications

***

### Summarizer: `llama3.2:3b`

**Used by**: Quick summaries, metadata extraction

**Strengths**:

* Fastest inference (\~1-2s first token)
* Lowest memory footprint (2GB)
* Good at extracting key points

**Weaknesses**:

* Limited context window (4K tokens)
* Less coherent on complex topics
* Not suitable for multi-turn chat

**When to use**:

* Summarizing scraped profile data
* Extracting metadata from long text
* Quick content previews
* Notifications and alerts

***

### Embeddings: `bge-m3:latest`

**Used by**: AnythingLLM document embeddings, RAG

**Strengths**:

* State-of-the-art multilingual embeddings
* Fast batch processing
* Supports 8K token context

**When to use**:

* Ingesting documents into AnythingLLM workspaces
* Semantic search over creator content
* Memory recall ("What did I post last week?")

**Not used for**: Text generation (embeddings-only model)

***

## Hardware Constraints

**Current server**: IONOS dedicated VPS, CPU-only (no GPU)

**Performance characteristics**:

* Models up to 7B: Acceptable performance (2-5s first token)
* Models 8B+: Stall on CPU-only (dolphin3:8b takes 45s+)
* Recommended limit: 7B quantized models (q4\_K\_M)

**Future upgrade path**:

* GPU VPS: Enables dolphin3:8b and larger models
* Alternative: Stick with Qwen-2.5 and optimize Action Runner flows

**Resource limits**:

* \~10GB RAM available for Ollama (after OS + other services)
* \~33 concurrent Stagehand browser sessions (300MB each)
* FFmpeg clip generation is the real bottleneck (\~30s CPU per clip)

***

## Model Selection Logic

```mermaid theme={null}
graph TD
    A[User request] --> B{Request type?}
    B -->|Chat/Q&A| C[qwen-2.5]
    B -->|NSFW caption| D[dolphin-mistral:7b]
    B -->|Taxonomy tag| E[scout-fast-tag]
    E -->|Fails| F[phi-3.5]
    B -->|Summary| G[llama3.2:3b]
    B -->|Multi-step flow| H[dolphin3:8b]
    H -->|CPU stall| C
    B -->|Document ingest| I[bge-m3]
```

***

## Configuration

### Environment Variables

```bash theme={null}
OLLAMA_URL=http://127.0.0.1:11434
OLLAMA_MODEL=qwen-2.5:latest  # Default for MCP server
STAGEHAND_MODEL=ollama/qwen-2.5  # For Stagehand vision tasks
```

### AnythingLLM Workspace Settings

**Administrator workspace**:

* LLM: `qwen-2.5:latest`
* Embeddings: `bge-m3:latest`
* Agent mode: Enabled
* Temperature: 0.7
* Max tokens: 4096

**Per-user workspaces** (future):

* Same settings, isolated context
* Potential routing to dolphin3:8b after GPU upgrade

***

## Related

* [MCP Servers](/ai/mcp-servers) — The `ollama` MCP server exposes these models
* [Action Runner](/ai/action-runner) — Uses models via `ollama` step type
* [Taxonomy System](/ai/taxonomy-system) — Uses scout-fast-tag for classification
* **Endpoint**: `POST /api/captions/generate` — Uses dolphin-mistral:7b