No fake “fastest” claim
Local AI performance changes dramatically with GPU, CPU, memory speed, model architecture, quantization, context, prompt length, driver, and thermal limits. Local Idea Studio will not call itself the world’s fastest runner without a broad, reproducible comparison.
A benchmark is trustworthy only when another person can repeat it. Results without hardware, model, settings, prompt, and token counts are marketing—not evidence.
Required test record
- Exact Local Idea Studio and runtime version
- CPU, GPU, VRAM, system RAM, operating system, and driver
- Exact GGUF repository, filename, quantization, and checksum
- Backend and actual GPU layers offloaded
- Context size, batch size, threads, and flash-attention state
- Prompt-processing speed, generation speed, generated tokens, and elapsed time
- Cold load time separately from warm generation
Current verified behavior
| Setting | Expected behavior | How to verify |
|---|---|---|
| Auto + NVIDIA | Prefer CUDA when supported | Runtime info shows cuda |
| Auto + other supported GPU | Use Vulkan when available | Runtime info shows vulkan |
| CPU only | No model layers offloaded to GPU | Runtime info shows cpu |
| GPU layers: all | Request maximum GPU offload | Loaded model reports actual offloaded layers |
Why results are not posted yet
The public release is new, and one developer computer is not a representative benchmark suite. Measured tables will be added only after the same test can be repeated across CPU-only, NVIDIA CUDA, and Vulkan hardware. Until then, Local Idea Studio displays live tokens per second in the app so users can measure their own exact system.
Run a useful comparison
- Use one model file and one fixed prompt.
- Start with the same context and generation settings.
- Generate at least 128 tokens after a warm-up run.
- Repeat three times on CPU, then CUDA or Vulkan.
- Report the median—not only the fastest run.
- Watch temperature and clock speed to identify throttling.
Share complete results through the GitHub issue tracker so future tables can link to evidence.