Re:Gemini 4 Neon
by jaykru on Wednesday November 18, 2026 @05:12PM
(#50552987)
Carbon 4 has been incredible for my use, tuning a local inference engine based on tinygrad running on Tenstorrent hardware. I had a barely functional prototype I built with Claude that petered out at an unusable level of performance. Astra and Sol didn't make much progress when I pointed them at it. Carbon 4 more than 10x'd the prefill and doubled the decode performance on Qwen 3.8 27B without breaking quality or the generality of the stack. Very nice to see competition at the frontier again.
[ Parent
| on HN ]
Re:Gemini 4 Neon
by atomashevic on Wednesday November 18, 2026 @05:26PM
(#50553182)
Just curious, what TT hardware do you have? How many tps did you achieve with Qwen 3.8 27B?
I'm running 2x p150a's right now. I have them hooked up to a mini-ITX PC with a single PCIe slot via a quad x1 switch (figuring out power to wire up the other two cards I own). My best numbers so far are around 2500t/s prefill about 52t/s decode with DFlash2. All of this is vibecoded besides the original design sketch which was maybe a 3 line description of what tweaks to make to tinygrad and which parts of the existing TT stack to use (almost nothing).
Re:Gemini 4 Neon
by henrymerrilees on Wednesday November 18, 2026 @10:17PM
(#50555614)
This sounds like a blast!
So much fun! 3D printed a little rack for it too. I'm just dreading figuring out the household electrical stuff as I scale it up ;)
Re:Gemini 4 Neon
by blueblisters on Wednesday November 18, 2026 @11:03PM
(#50555940)
A plausible explanation here is that Ant and OAI are deliberately sandbagging on ML / GPU optimization tasks to preserve their lead while Google doesn’t give a fuck.
Argon is incredible at AI research and its responses are grokkable
It's also really good at my fuzzier math/research questions :shrug: more insightful than fable imo. The nerfing is real (indeed Ant at least admit to it) but I don't really feel like it's a useful fact to have in mind when evaluating the products I pay for.