Oh, there's an AI party! O,,O) Let me say hello and share some quick test results of the Gemma4 model.
The test environment was an hp zbook ultra g1a (approximately 5060 level raw performance, 128GB unified memory, Ollama).
First, after loading the model, the initial response speed to the question "Who are you?" is as follows:
Gemma4:e2b (2.3B) : approximately 3.8 seconds
Gemma4:e4b (4.5B) : approximately 7.0 seconds
Gemma4:26b-A4b (26B) : approximately 15.9 seconds
Gemma4:31B (31B) : approximately 65.7 seconds
The speed for a simple syllogism evaluation is as follows. ("Evaluate the following syllogism. 'All humans eventually die.' 'Mr. Kim next door is human.' 'Mr. Kim next door will eventually die too.'")
Gemma4:e2b (2.3B) : approximately 8~14 seconds
Gemma4:e4b (4.5B) : approximately 15~17 seconds
Gemma4:26b-A4b (26B) : approximately 34 seconds
Gemma4:31B (31B) : approximately 157 seconds
And more importantly than speed, the quality of the output results is truly remarkable.
Even the smallest model, e2b, supports multilingual text. The results output from text-to-text are very satisfactory.
With previous versions like Gemma2 or Gemma3, the base models were embarrassing to use without fine-tuning.
With Gemma4, it seems like decent performance can be achieved from the base model with just good prompts.
I'm excited because I expect a certain level of performance in document classification in the text-to-text field that I mainly work with.
When LLM language models first came out, larger models showed better performance.
Language models for edge devices are also showing considerable capability (subjectively, they seem better than ChatGPT version 4).
Some Local AI is currently operating via MCP. I'm looking forward to seeing programs (or package services) that demonstrate impressive performance with Local AI arriving soon. O..O)b
