Ollama Local Models

SkopaqTrader supports Ollama as a local LLM fallback. When cloud API keys are unavailable or you want to run analysis offline, Ollama provides a self-hosted alternative.

How It Works

Ollama is the last option in each agent's fallback chain. If all cloud providers fail (missing keys, rate limits, errors), the system automatically falls back to a local Ollama model.

# skopaq/llm/model_tier.py — fallback chain per role
"market_analyst":  [("google", "gemini-3.8-flash"), ("ollama", "auto")]
"social_analyst":  [("openrouter", "x-ai/grok-4.6"), ..., ("ollama", "auto")]
"trader":          [("google", "gemini-3.8-flash"), ("ollama", "auto")]

Judge roles excluded

The research_manager and risk_manager roles do NOT have Ollama fallback. These judge roles require the strongest reasoning quality, so they only fall back from Claude Opus to Gemini -- never to a local model.

Enabling Ollama

Ollama is opt-in. Set the environment variable:

SKOPAQ_OLLAMA_ENABLED=true

Without this, Ollama is never checked, even if it is running locally.

Setup

Step 1: Install Ollama

# macOS
brew install ollama

# Linux
curl -fsSL https://ollama.com/install.sh | sh

Step 2: Pull a Model

ollama pull mistral        # Default fallback (7B, fast)
ollama pull llama3.1:8b    # Good general purpose
ollama pull qwen2.5:14b    # Stronger reasoning, needs more RAM

Step 3: Start the Server

ollama serve

Ollama runs at http://localhost:11434 by default.

Step 4: Configure SkopaqTrader

# Required: opt-in to Ollama
SKOPAQ_OLLAMA_ENABLED=true

# Optional: custom base URL
SKOPAQ_OLLAMA_BASE_URL=http://localhost:11434

# Optional: specify model (otherwise auto-detected)
SKOPAQ_OLLAMA_MODEL=mistral

Model Auto-Detection

When SKOPAQ_OLLAMA_MODEL is not set (or set to auto), the system auto-detects the best available model:

def _get_ollama_model() -> str:
    # 1. Check SKOPAQ_OLLAMA_MODEL env var
    # 2. Query Ollama API for installed models
    # 3. Use the first available model
    # 4. Default to "mistral" if nothing found

The detection queries http://localhost:11434/api/tags and picks the first installed model.

Availability Check

On first use, the system pings the Ollama API to check if it is running:

def _is_ollama_available() -> bool:
    # 1. Check SKOPAQ_OLLAMA_ENABLED is set
    # 2. HTTP GET to /api/tags
    # 3. Cache the result for the session

The result is cached -- the check only happens once per session.

LangChain Integration

Ollama models are created via langchain-ollama:

from langchain_ollama import ChatOllama

llm = ChatOllama(
    model="mistral",
    base_url="http://localhost:11434",
    temperature=0.3,
)

The ChatOllama class implements the same BaseChatModel interface as all other providers, so it works seamlessly with the LangGraph pipeline.

When Ollama Activates

The fallback chain is evaluated left to right for each agent role:

market_analyst:
  1. Try Google (Gemini 3.8 Flash)
     → GOOGLE_API_KEY set? → Use it
  2. Try Ollama (auto)
     → SKOPAQ_OLLAMA_ENABLED=true AND Ollama running? → Use it
  3. No provider available → Error

Ollama will activate if:

All cloud API keys for that role's chain are missing or empty
SKOPAQ_OLLAMA_ENABLED=true is set
Ollama is running and responds to the health check

Supported Models

Any model available through Ollama works. Recommended models for trading analysis:

Model	Size	RAM	Quality	Speed
`mistral`	7B	8 GB	Good	Fast
`llama3.1:8b`	8B	8 GB	Good	Fast
`qwen2.5:14b`	14B	16 GB	Better	Medium
`llama3.1:70b`	70B	48 GB	Best	Slow
`deepseek-r1:14b`	14B	16 GB	Strong reasoning	Medium

Quality trade-off

Local models are significantly less capable than cloud models (Claude Opus, Gemini) for financial analysis. Use Ollama as a fallback for experimentation, not for production trading decisions.

Configuration Reference

Variable	Default	Description
`SKOPAQ_OLLAMA_ENABLED`	`false`	Opt-in flag (must be `true`, `1`, or `yes`)
`SKOPAQ_OLLAMA_BASE_URL`	`http://localhost:11434`	Ollama server URL
`SKOPAQ_OLLAMA_MODEL`	`auto`	Model name or `auto` for auto-detection

File Reference

File	Purpose
`skopaq/llm/model_tier.py`	Fallback chains, Ollama detection, model creation
`tests/unit/llm/test_ollama.py`	Unit tests for Ollama integration