Poor performance with local models

Description

I am using LM Studio with a local model, and responding in OpenCode takes much much longer (I would say 20x) than directly in LM Studio. I have also already tried with Ollama and the performance was poor as well. Is there a special reason for that, or could that be improved?

OpenCode version

1.0.67

Steps to reproduce

Use Local model with LM Studio (Ollama the same)
Wait for a response in OpenCode and compare to LM Studio directly

Screenshot and/or share link

No response

Operating System

Windows 11

Terminal

Windows Terminal

This issue might be a duplicate of existing issues. Please check:

Poor Performance and High Resource Usage with Ollama Models in opencode #4182: Poor Performance and High Resource Usage with Ollama Models - reports similar performance degradation when using Ollama through opencode (0.5 tokens/second vs 16 tokens/second directly)
OpenCode v1.0.25 Hangs Indefinitely with LM Studio + Qwen Models Due to Empty tool_calls Array #4255: OpenCode v1.0.25 Hangs Indefinitely with LM Studio + Qwen Models - detailed investigation of LM Studio performance issues, identifies root cause related to tool_calls handling
App is calling the wrong endpoint for the Ollama provider when using local models #4341: App is calling the wrong endpoint for the Ollama provider - reports incorrect API endpoint usage that could affect performance

The issues show a pattern of local model performance problems across both LM Studio and Ollama providers. Issue #4255 provides particularly detailed analysis suggesting the root cause may be related to empty tool_calls array handling.

Feel free to ignore if none of these address your specific case.

@JimmyBenHur

@JimmyBenHur I assume this is because opencode sends a lot more context right? tools + system prompt = ~10k tokens, add in your AGENTS.md files or mcp servers and it is even more

Definitely makes sense that it's slower because of the context, but this much? The issue seems to be similar to #4182, but I have no AGENTS.md and no mcp servers, I wouldn't except it to be that much slower. I have about 12 tokens per second in LM Studio and I would say less than half a token in OpenCode, when using Qwen3 Coder. Isn't there something to do to improve performance?

I just don't know how we could possibly be causing an issue for you? We send an http request and the results stream in

Here is a good way to test it, make an opencode.json file and add:

{
  "tools": {
     "*": false
  }
}

That should disable all tools, and then see if it is any faster

Yeah, it is actually much faster without the tools. Thing is I need the tools...

@JimmyBenHur

@JimmyBenHur yes the tools are the whole thing as you know, but that's proof that the context is slowing down your local model, it isn't an opencode issue right?

Looks like it isn't...

To stay organized issues are automatically closed after 90 days of no activity. If the issue is still relevant please open a new one.

Navigation Menu

Repository navigation

Poor performance with local models #4385

Description

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Poor performance with local models #4385

Description

Description

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Activity

github-actions commented on Nov 16, 2025

rekram1-node commented on Nov 16, 2025

JimmyBenHur commented on Nov 16, 2025

rekram1-node commented on Nov 16, 2025

JimmyBenHur commented on Nov 16, 2025

rekram1-node commented on Nov 17, 2025

akemmanuel commented on Jan 7, 2026

thdxr commented on Mar 25, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions