Qwen3 small model actual test: from 4B to 30B, which one can use MCP to communicate smoothly with Obsidian?

I heard it was released last night qwen3 The model's Agent and code capabilities have been optimized to strengthen support for MCP.

Qwen3: Think deeply and move quickly
https://qwenlm.github.io/zh/blog/qwen3/

This sentence in the introduction

The small MoE model Qwen3-30B-A3B has 10% of the number of activated parameters of QwQ-32B, and performs better. 
`A small model like Qwen3-4B can also match the performance of Qwen2.5-72B-Instruct`.

I was very excited, so I went home after get off work. nas server use Ollama Pull model deployed, use cherry studio, enable obsidian-mcp, started testing, but the test results slapped me in the face.

Test content:

1. 1. Check my obsidian knowledge base for changes in the last day, and the model will answer randomly

The model cannot hit the tool.

1. 1. Use obsidian's mcp obsidian_get_recent_changes tool to query the changes in my knowledge base in the last day

I prompted the name of the tool, but the model still gave a random answer.

qwen3 model

Model evaluation item description

Obsidian-MCP

Obsidian-MCP is commonly used for the following tasks:

• • Semantic retrieval and summarization of log/note content (embedding + question-answering)
• • Self-dialogue (multi-turn historical context)
• • Context-based "thinking enhancement" such as task suggestions and card associations
• • Memory callback of private knowledge base (streamable / SSE mode long connection)
• • Local embedding + lightweight reasoning, no reliance on public network LLM

JSON search to obtain the content of periodic notes. Get the list of recent periodic notes. Get the most recently modified files.

Test whether Qwen3-4B's capabilities match the above requirements

Test whether Qwen3-8B's capabilities match the above requirements

Use cherryStudio to test qwen3:8b. It is possible to call the tool, but the answer is hallucinatory and the name of the returned note is changed.

Qwen3-4B-Local Model + Obsidian-MCP's `Local Q& A`.md
The answer became

01Project / Blog/draft/Qwen3-4B-Local Model + Obsidian-MCP's `Local Issues`.md

At this time, the notes are used git sync The advantage comes out. When you use mcp locally to organize your notes, if an error occurs, you can roll back to the last submitted version at any time!

Test whether Qwen3-14B's capabilities match the above requirements

Using openrouter qwen3:14b Model testing

I think it is a limitation of openRouter's own deployment. qwen3-demo:

https://huggingface.co/spaces/Qwen/Qwen3-Demo

After testing, the same text can be summarized normally, and 128k tokens are enough. It seems that 8B, 14B, and 32B can still be used locally.

in conclusion

The knowledge base interaction test using Qwen3 and Obsidian-MCP concluded that:

Version 4B : Quantization compression leads to aphasia

• • The tool call capability is completely lost, facing clear obsidian_get_recent_changes Indifferent to instructions
• • The token capacity is 32K, so long sessions may be difficult to process completely

Version 8B : Seemingly useful but actually dangerous

• • Although the tool call can be recognized, the returned file path has a high error rate;
• • Appears when the content is summarized Hallucination Rewrite, the note name will be modified;
• • If the MCP API is accidentally deleted and there is no git backup, it will be more dangerous

Version 14B+ : Really Fragrance Warning

• • 128K token capacity perfectly adapts to the knowledge base scenario, and accurately calls the Obsidian API during testing
• • However, local deployment requires 16G video memory, which is prohibitive for most NAS users

Before my 16G graphics card arrives, I have to pay attention to privacy protection. I first use the cloud-based large model + MCP to read the non-sensitive data directory as the context for questions and answers.

After all, to be a technology master, you must understand Finding the optimal solution within realistic constraints.