Nick

[Nick @sameobserver]

@opencode Hy3 is debited at regular rates, without discount

[Nick @sameobserver]

@NousResearch Please fix the hermes desktop update for Windows. The hell doesn't work.

[Nick @sameobserver]

@datacurve And why do you make such crazy, incomprehensible graphics, and even in a video.

[Nick @sameobserver]

@jun_song Not only is it more expensive, but it is also significantly dumber

[Nick @sameobserver]

@claudeai Why the do you need it at such a price, when there are models smarter and cheaper by an order of magnitude?

[Nick @sameobserver]

@Polymarket Why write JUST IN: before every tweet? What the? Do you give a fuck?

[Nick @sameobserver]

@0xSero I think it's close to qwen 35b in terms of speed, but much dumber

[Nick @sameobserver]

@NVIDIAAI Why is it needed?

[Nick @sameobserver]

@finkd https://t.co/JnOXSrONhW

[Nick @sameobserver]

A real revolution in local models will be made by the one that, with quantization of 4, will occupy about 25 GB and have 1-2 billion active parameters. This will allow you to run it on almost any PC with 32 GB of RAM, regardless of the presence of a graphics card.

[Nick @sameobserver]

@alexandr_wang https://t.co/JnOXSrONhW

[Nick @sameobserver]

A real revolution in local models will be made by the one that, with quantization of 4, will occupy about 25 GB and have 1-2 billion active parameters. This will allow you to run it on almost any PC with 32 GB of RAM, regardless of the presence of a graphics card.

[Nick @sameobserver]

And with a modern conventional consumer graphics card, it will issue more than 100 tokens per second

[Nick @sameobserver]

A real revolution in local models will be made by the one that, with quantization of 4, will occupy about 25 GB and have 1-2 billion active parameters. This will allow you to run it on almost any PC with 32 GB of RAM, regardless of the presence of a graphics card.

[Nick @sameobserver]

@Tech2Wild Even 1x5090+ ram will give comparable or greater speed than 2 dgx. Calculate the average memory speed to estimate the average speed

[Nick @sameobserver]

@tonysimons_ Please make the target ratio in tokens, not just the coefficient. So that you can compress the context not only to 0.2, but also to 50k

[Nick @sameobserver]

@Jason Singularity is when they choose us, not us

[Nick @sameobserver]

@imnotchalk I just started testing Luna when the new Deepseek came out)

[Nick @sameobserver]

@reach_vb Add the new Deepseek V4 Flash to the chart and the picture changes dramatically

[Nick @sameobserver]

It's just that Deepseek V4 Flash 1031 will be the first in all AI rankings

[Elshayib @elshayib_]

Deepseek V4 flash 0731, what's up with the naming tho?
Previous was called "Preview" which makes sense but "0731" that's odd.

[Nick @sameobserver]

@Mylovanov 44 tons is about 2%, not half. Let him learn to count, and it would not hurt you to

[Nick @sameobserver]

@thorstenball The difference, in fact, is only in either writing or pressing the plan button. And yet, in working mode, he can begin to change something at his discretion, which is not good