Local LLM models working well with maverick

[WSINTRA]

I'm using ollama with open-web-ui, got the mcp server up and running with mcpo, but seem to be having hit and miss with tool calling, been trying with gpt-oss and qwen3-30 with mixed results.
I'm also not really up to scratch on these kinds of tools, I'm mostly playing around so my son can have a go. I shared the creators blog so he can learn about the technical goings on, but my question. What open models are good for calling maverick and what kinds of prompts are good for triggering the tools, direct calls, like stock_price(ticker) or more conversational stuff like "stock price for apple today"
How big of a difference does it make using open source models vs openAI/Anthropic SOA models ?

[wshobson]

Hey @WSINTRA! Cool to see you and your son exploring local LLMs with MaverickMCP!

Quick Model Recommendations

For Ollama, these work well with MaverickMCP's tools:

• Mixtral 8x7B - My go-to for balanced performance
• DeepSeek-Coder 33B - Surprisingly good at tool calls
• Qwen 2.5 32B - Your Qwen3-30 choice is solid!
• Llama 3.1 8B - Great lighter option

Prompting Tips That Actually Work

Direct is better with local models. Skip the conversational style:

• ✅ "Get AAPL technical analysis with RSI and MACD"
• ❌ "Can you tell me about Apple's indicators?"

I just added 12 test examples to the README today - they're battle-tested and work great with local models. Start with these:

Easy mode:
"Get AAPL stock data for the last 30 days"

Challenge mode:
"Research NVDA including fundamentals, technicals, and competitive position"

The Reality Check: Open vs Commercial

Local models get you ~75% of the way there vs commercial models' 95%+. But honestly? For most stock analysis tasks, that 75% is plenty. Local models are faster (no network lag), private, and free to
run.

Main difference: Commercial models handle ambiguity better. Local models need clearer instructions but absolutely get the job done.

Pro Tips for Your Setup

1. Keep temperature low (0.1-0.3) for consistent tool calling
2. Add this to your system prompt: "You have MaverickMCP tools. Use them directly when asked about stocks."
3. The new 7-256x speedup in parallel research works great with local models too!

Your son's learning with real tools, that's awesome! Let me know which models work best for you both. Always interested in real-world feedback from the community.

Happy hacking!