Poor Performance and High Resource Usage with Ollama Models in opencode #4182

Description

Environment:

• OS: Windows 10
• GPU: NVIDIA GeForce RTX 3060 Mobile (6 GB)
• Installation Method: Installed via choco install opencode
• Model Used: hf.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF:Q4_K_M

Description:
When using Ollama models through opencode, the performance is significantly degraded compared to using the same model directly with Ollama. Specifically:

• Token Generation Speed:

  • Ollama (direct): ~16 tokens/second
  • opencode: ~0.5 tokens/second or much slower

• Resource Usage:

  • CPU and GPU usage are much higher than expected when running the same model via opencode, even though the performance is much worse.

OpenCode version

opencode v1.0.55

Steps to reproduce

1. Install opencode via choco install opencode
2. Run a large language model (hf.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF:Q4_K_M) using opencode
3. Compare token generation speed and resource usage with direct Ollama usage
4. Observe significantly lower performance and higher resource consumption

Screenshot and/or share link

Even with the addition of a request from /opencode/src/session/prompt/qwen.txt, ollama gives 12 tokens per second, while opencode is around 0.5 or less.

Is there a way to measure token speed in opencode?

Operating System

Windows 10

Terminal

PowerShell