Testing Qwen3 Models for Local Agents with MCP - From 0.6B to 235B

Explore the capabilities of Alibaba's Qwen3 language models in this comprehensive video tutorial where Chris Hay tests various parameter sizes from the tiny 0.6B to the massive 235B model. Learn how these models perform with MCP (Model Control Protocol) for agentic scenarios using Ollama. The video demonstrates each model's capabilities through joke generation, tool calling with MCP CLI, and comparison with Llama 3.3. Discover the surprising capabilities of the smallest 0.6B model, examine the performance of the 30B mixture-of-experts architecture, and see complex telnet command execution and coding tasks with the larger models. The tutorial includes timestamps for easy navigation through different model evaluations and provides links to GitHub repositories for MCP CLI, MCP telnet client, and the CHUK protocol server for those wanting to replicate the experiments.

Syllabus

in this video chris looks at the new qwen3 models from the smallest qwen3 0.6B parameter model all the way upto the qwen3:235B model. Is a 600 million parameter model actually usable? Which model should use you use for local agents including the new 30B mixture of expert model
00:00 - introduction
01:35 - running qwen3 with ollama
02:49 - one joke with many models
06:40 - calling mcp servers with qwen3:0.6B and mcp cli
12:07 - comparing llama3.3 with qwen0.6B
13:49 - calling multiple tools at once
15:58 - qwen3 1.7B
17:05 - qwen3:8B
19:25 - qwen3:30B mixture of expert model
21:23 - can qwen3:0.6B telnet
23:47 - complex telnet commands with qwen3:30B
33:30 - qwen 235B and coding