Repository navigation
Local llama.cpp models are extremely slow when session title generation is enabled #33126
Description
Description
Environment
- OpenCode: 1.17.8
- Provider: llama.cpp (OpenAI-compatible)
- Model: Qwen3-14B-Q4_K_M
- OS: macOS (Apple Silicon)
- OpenCode installed via Homebrew
Problem
When using a local llama.cpp model, OpenCode is significantly slower than direct API calls.
Direct llama.cpp API
time curl -s http://127.0.0.1:8080/v1/chat/completions
-H "Content-Type: application/json"
-d '{"model":"q14","messages":[{"role":"user","content":"2+2=?"}}'
Result:
~2.5 seconds
OpenCode
time opencode run "2+2=?" --model llamacpp/q14
Result:
~22 seconds
OpenCode with fixed title
time opencode run "2+2=?" --model llamacpp/q14 --title test
Result:
~4.6 to 5.1 seconds
Additional findings
Debug logs show:
agent=title small=true
agent=build small=false
When --title is not provided, OpenCode appears to invoke an additional title-generation pipeline.
The performance difference is much larger than expected:
Direct API: ~2.5s
OpenCode + fixed title: ~5s
OpenCode auto title: ~22s
The title-generation path adds roughly 15~18 seconds.
Separate issue
I also found a large startup delay caused by external skill scanning.
Before disabling external skills:
init count=235
After disabling:
init count=1
Runtime improved from:
176s -> 29s
There may be an opportunity to lazy-load skills or avoid scanning duplicate skills.
Expected behavior
- Session title generation should not add 15~18 seconds for local models.
- Title generation should ideally use a lightweight path or be disabled automatically for local providers.
- Duplicate skill scanning should not significantly impact startup latency.
opencode --version
1.17.8
brew list opencode
/opt/homebrew/Cellar/opencode/1.17.8/bin/opencode
Plugins
No response
OpenCode version
No response
Steps to reproduce
No response
Screenshot and/or share link
No response
Operating System
No response
Terminal
No response
Activity
Necmttn commentedon Jun 21, 2026
This needs a latency breakdown per session start. Split timings for config load, external skill scan, first model call, title generation, and final response. Then local-model users can see whether the 20s delay came from title generation, skill inventory, or provider latency.
The observed
--title testspeedup is strong evidence title generation belongs behind a skip/defer policy. A regression can assert local llama.cpp with explicit title avoids the title agent path entirely.Generated with ax.
kevincox commentedon Jul 2, 2026
I also don't have hard numbers, but I notice that title generation usually generates thousands of tokens which seems very overkill for a title. Maybe reasoning is being enabled or something for title generation which is probably unnecessary? Ticket #25456 seems to think so. I agree that this is probably undesirable by default.
I wouldn't mind nearly as much if it was just ~550 prompt tokens and ~20 output tokens. But 2k output tokens can take a long time on some models or providers. The actual output surely depends on the model but it seems very common for them to be chatty.
I also noticed that setting your own title while it is running:
Getting more off-topic I wonder if title generation could also be handled differently. Like what if there was a tool to set the title. The system prompt tells the model to set the title near the start and whenever there are significant changes in direction. The tool starts returning errors once the user has explicitly edited the title and telling the model to stop calling it. (Probably a bad idea, IDK if I want my sessions changing title as they progress, but worth considering).
github-actions commentedon Sep 1, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.