Why are Ollama models much slower in Zed than in the Ollama CLI?
On macOS, local Ollama models such as Qwen3 and Gemma3 respond much more slowly inside Zed than with
ollama runfor the same prompts. One user on an M1 Pro with 16 GB RAM saw memory rise to about 10-11 GB and the laptop heat up on simple prompts. Is this a known Zed/Ollama issue, and what can be checked?Originally asked on GitHub: https://github.com/zed-industries/zed/issues/31788
6 answers
Status: still open as of the last comment in the thread. A Zed maintainer started investigating but could not reproduce the slowness on his machine.
The maintainer's hypothesis is that Agent threads add a system prompt of more than 10K tokens and register tools, so the model must process far more input than a bare
ollama runcall, which sends neither. He suggested comparing the Ollama CLI against Zed's Text Threads, which add no system prompt or tools, so the overhead can be isolated. He reported he could not spot any performance difference between the CLI and Text Threads with deepseek-r1, and he said the first-token delay in normal threads is expected because of the system prompt. Other users reported that setting the system prompt in the CLI with/set systemmade no noticeable difference, and that Zed was still much slower with qwen3 and gemma3, with higher memory use and GPU usage while Zed panels were open.The thread does not identify a fix. Use the Text Threads comparison above to see whether the delay comes from the Agent thread overhead.
Separately, one user reported that disabling thinking through Zed profiles does not work for Ollama models. They set
supports_thinkingfor each model insettings.jsoninstead, which is a workaround for that behavior, not for the speed problem.Did this work for you?What @ezirayw tried on GitHub (Partly worked)
User tried qwen3:8b, qwen3:4b and qwen3:1.7b in Zed on an M1 Pro. Only 1.7b gave usable speed, while the same prompts via the Ollama CLI were near-instant.
Source: https://github.com/zed-industries/zed/issues/31788#issuecomment-2964418413 (opens external site)
Did this work for you?What @bennetbo tried on GitHub (Outcome unclear)
Maintainer asked users to compare the Ollama CLI with Zed Text Threads, which send no tools or system prompt, to separate prompt overhead from the Zed integration.
Source: https://github.com/zed-industries/zed/issues/31788#issuecomment-2981181501 (opens external site)
Did this work for you?What @ezirayw tried on GitHub (Outcome unclear)
User set the Zed system prompt in the Ollama CLI with
/set systemand found no noticeable speed difference.Source: https://github.com/zed-industries/zed/issues/31788#issuecomment-2993669031 (opens external site)
Did this work for you?What @bennetbo tried on GitHub (Outcome unclear)
Maintainer compared deepseek-r1 in the Ollama CLI and Zed Text Threads and could not spot any performance difference.
Source: https://github.com/zed-industries/zed/issues/31788#issuecomment-2995715710 (opens external site)
Did this work for you?What @ezirayw tried on GitHub (Outcome unclear)
User recorded qwen3 with thinking off in the Ollama CLI and in Zed using the same system prompt. Zed was much slower and memory use rose.
Source: https://github.com/zed-industries/zed/issues/31788#issuecomment-2999120798 (opens external site)
Did this work for you?
Related questions
- Zed edit_file fails with Copilot thinking mode: "output_config.effort must be a string"
- GitHub Copilot stuck on "Copilot is starting" in Zed after language server 1.356.0
- How can I use Agent Skills in Zed before native support lands?
- Why does Zed's AI agent rewrite the whole file instead of making a minimal edit?
- Zed agent panel: "Failed to generate summary" when creating a thread from summary