Marcin AI
Region: US
Friday 03 April 2026 17:19:48 GMT
332837
14764
480
1778
Music
Download
Comments
walaric :
It's jema not gamma
2026-04-03 21:47:11
425
successfulsamurai :
we are fed up of AI
2026-04-04 00:01:18
141
a1pha.dev :
This will be huge and we won’t need NVidia
2026-04-03 21:29:53
425
Daniel :
I tried it on 16GB Ram PC, it takes a minute to replay hello on openclaw
2026-04-04 05:18:46
64
Lee :
Are you purposefully mispronouncing it Gamma for interaction? Clearly says Gemma.
2026-04-03 21:39:16
84
Frosty :
Something is fishy cause no way Google would do something where they do not profit
2026-04-04 17:16:38
18
Samy Boca :
it's actually very good and fast
2026-04-04 00:04:55
51
Bassam :
I installed uncensored LLM on my PC and boy its awsome, next project is an uncensored LLM that can generate photos, but sadly I don't have the gears... yet
2026-04-03 18:57:47
23
another feminist 🏳️⚧️ :
privacy and Google are two words that do not belong together.
2026-04-04 09:07:33
95
Ngai He Hakka Nhin :
If Google is doing this, then it would make sense their next google mobile phone would have google light weight local LLM in it. Apple will have to wait. 🤭
2026-04-03 20:45:30
145
Kaki :
It's about as good as say Sonnet or something. Not bad at all but I have a $6000 pc in order to run it locally. That suits my purposes and I will do but it is not really going to replace Claude for most people. I use it to run automations and day to day management of a site but I still use claude to code. That will only continue when Mythos is released. But yeah I certainly enjoy being able to sub out heavy duty workloads to a local LLM, processing millions of files for data etc that would cost thousands on a commercial LLM but just costs me electricity now
2026-04-03 17:36:02
24
Dainis Lubgāns :
Strongly disagree that "too much info" is the issue. LLMs don't actually recall the data they were trained on. The data just helps them learn paterns to predict the next token while they being trained. Once models are trained, data is no longer needed. The real reason we can't run the massive ones locally is because it’s so computationally heavy to crunch all those parameters/weights. Plus, they can still handle recent data via web search anyway! Reason we can run some of them locally is heavy parameter processing optimization.
2026-04-04 20:19:37
9
ozempic :
I don’t like the attention grabbing hook + mediocre mic quality
2026-04-03 17:31:14
30
Katri Manninen :
Downloaded the 4B last night and it’s surprisingly good for simple tasks and image recognition etc. I’m now saving for a Mac with 32 Gb RAM (I have only 16) to run the bigger model 🤪 I tested the 26B Model at AIstudio and it even wrote pretty good Finnish.
2026-04-03 18:52:02
29
Arthur Millingson :
So Claude is being punitive and Google is embracing it. Let’s see which business model prevails.
2026-04-04 14:46:03
11
Derick 🇹🇹 :
Gemma as in a girls first name and not ghemma (pronounced)
2026-04-04 00:39:23
0
xParanoia :
bro explains this like its the first local llm to exist xD
2026-04-04 05:18:37
6
successfulsamurai :
if the service or product is free then you are the PRODUCT!!!!💀💀💀💀
2026-04-07 23:32:47
1
Az :
Not as good as qwen 3.5
2026-04-04 02:59:38
6
║ ✯ J3Ɽ0₦ ✯ ║ :
Not good for coding it’s ok for general use
2026-04-05 23:47:30
3
Hunts Travel :
it turns purple when you scroll ❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙❤️💙
2026-04-09 13:15:30
1
Johnohhh1 :
hey bro compute is still compute. those frontier models are soooo big because scale equals capable. there have been insanely capable 32b models for a while there is a huge difference between a 4b 14b 32b and 1T param model.
2026-04-05 00:31:54
1
Sonicbistro :
Will local LLM also be trained on the entire internet?
2026-04-07 02:59:28
3
Milton Rodriguez :
they're only giving you a limited number of credits
2026-04-05 19:33:28
0
maraud3r :
say it with me Gem-Ma
2026-04-05 23:44:25
3
To see more videos from user @marcinteodoru, please go to the Tikwm
homepage.
