The bigger the model, the more capable GPU/NPU/CPU it needs in addition to the memory bandwidth.
The RTX PRO 6000 videos shows when he's running Qwen2.5 Coder 32B FP16 with the size of 61GB, even M4 Max with memory bandwidth of 500GB/second only getting 7.63 tokens/second while the RTX PRO 6000 still getting good speed with 23 tokens/second. Ryzen AI Max+ 395 he uses Qwen2.5 Coder 32B q4_k_m which is only 20GB but only getting 10.8 tokens/second. This 395 CPU is very capable but it's limited by the memory bandwidth.

Gemini now can generate image with accurate long text! I tried and there's not even 1 wrong alphabet. But the image is not as highly detailed as generated by Flux. And the text is simple without much details


Flux generate a lot higher details than Gemini but it still have issue to make accurate text, and can't generate all text if it's too long. I have to generate tens of images to get to this one image that looks good and first part of text is accurate.


Just for fun I try create image with added prompt for Studio Ghibli artistic styles. First is Gemini and second is Flux

Yup, if need something fast and presentable, Gemini is good enough just like camera phone.

I watched this video of data center with clusters of GPU, it is hundreds or thousands time more powerful than any PC's GPU. But the image created by Gemini not up to the level of Flux, actually Google make hard restriction of Gemini ability so that they can save the processing power.

I don't understand. I been running ollama without docker from the start. Or you mean WebUI? Didn't see any instructions on how to do this.

Just updated ollama and I understand what you mean now. No need to use WebUI anymore.
The setting is so simple, can't even see it's token per second. But looks like more faster than WebUI and more accurate too.

Qwen image sounds promising. Can do long text. Haven't see yet if can use with Comfy UI

QWEN3 the latest high performance LLM


But the size is bigger than 32GB for quantize 8bit and bigger than 16GB for quantize 4bit. LM Studio support splitting into 2 GPUs for out of memory issue but the tokens per second is a lot slower than Apple's M chip


interestingly there's a red bar that flashes for a second and the name is blurred out. Could that be the AMD AI Max+ 395? He did say he couldn't share AMD result until next week when someone ask, maybe still under embargo.

Been a while since I last play with LLM. I tried GPT-OSS:20B that came with Ollama app. It is thinking model but were able to came to conclusion at less than 2 seconds, much quicker than deepseek-r1 that think for very long for simple calculation and took almost 37 seconds.


Asking this question on deepseek-r1, GPT-OSS:20B and gemma3:12b


deepseek-r1 understand the story wrongly by saying Colin reveal each box content and that also led to wrong conclusion



gemma3:12b doesn't show the thinking, it manage to understand the story correctly and also correctly decide the decision Bob should take



GPT-OSS:20B goes PhD level of answer after thinking for 3.1 seconds

Latest Ollama also improved on the character recognition. All the character correctly presented. But Ollama app still can't add picture, have to use Open WebUI

The photo shows up to 128GB RAM but there's no option to configure the machine 🤷

Meanwhile I'm having fun with Google's Nano Banana. It's as good as running local generative such as Flux

I add this photo


And ask it to create a realistic photo based on that image.


I ask it to change the background to be on rooftop and with some slight digital effect on the side of the image


I ask it to change it to daytime


Make explosion effect on the background similar to Hollywood big budget movie (it block when I add Hollywood in the prompt, I then change it to big budget)


Add bold heading on the bottom, make the word appear to be made on neon light bulb:
lowyat.net
malaysia's largest online community


It's ability to be consistent in each new photo generation is amazing.

Qwen release their image_edit_2509 recently. It supposed to be able to edit image with prompt and remain consistent. I tried the lowyat picture and ask to remove the explosion. It did well


Using Nano Banana


Qwen able to output very close to Nano Banana, but it didn't remove the flare on the middle of the two guys shirt as cleanly as Nano Banana

I tried ask Qwen to create realistic photo from the original design


This post has been edited by xxboxx: Sep 28 2025, 11:14 PM

This is using ComfyUI for image generation.

deepseek-ocr was released yesterday and it supposed to be able to see the image and ran OCR but when I ran it, it repeating "x=0" until I close the app. Already updated to version 0.13.0, don't know what is the issue. Maybe still have some bugs

Yeah, at first it seems better with it's app, use less resource and more faster but now it seems more worse.
Also very annoying it add cloud in the choice for model.

The VRAM requirement is getting even more higher