Model Selection

import { Aside } from ‘@astrojs/starlight/components’;

Tool calling is a hard requirement for the ReAct agent — not all models support it correctly. This guide helps you choose the right one.

By available RAM

Available RAM	Model	Command	RAM used	Notes
8–16 GB	qwen3:8b	ollama pull qwen3:8b	~5 GB	Fast, minimum viable quality
16 GB	qwen3:14b ⭐	ollama pull qwen3:14b	~9 GB	Recommended — best quality/RAM ratio
32 GB	qwen3:14b or qwen3:30b	ollama pull qwen3:30b	~19 GB	30b = MoE, faster, 256K context
48 GB+	qwen3:32b	ollama pull qwen3:32b	~20 GB	Best quality, dense
48 GB+ (long docs)	qwen3:30b	ollama pull qwen3:30b	~19 GB	MoE, 256K context — ideal for large evidence files

After pulling, set OLLAMA_MODEL in your .env:

OLLAMA_MODEL=qwen3:14b   # or whichever you pulled

What to avoid

Model	Why
qwen3.5:*	Tool calling broken in Ollama (March 2026) — format mismatch causes HTTP 500 on every tool call (#14493)
qwen2.5	Superseded by Qwen3 across all benchmarks
Models without tool calling	The ReAct agent cannot function

All Qwen3 models listed above support multilingual output including Spanish and English.

Cloud alternatives (no Ollama)

If you don’t want to install Ollama or need maximum quality:

# OpenAI
LLM_PROVIDER=openai
OPENAI_MODEL=gpt-4o
OPENAI_API_KEY=sk-...

# Anthropic
LLM_PROVIDER=anthropic
ANTHROPIC_MODEL=claude-3-5-sonnet-20241022
ANTHROPIC_API_KEY=sk-ant-...

With cloud LLMs, your investigation queries are sent to the provider. If data privacy is important, use local Ollama.

How to change the model

# 1. Pull the new model
ollama pull qwen3:32b

# 2. Update .env
OLLAMA_MODEL=qwen3:32b

# 3. Restart the agent
osint-agent

No service restart needed — the agent reads the model from .env on each startup.