Lex Savage

I'm just messing around I like AI. I'm not posting anything so don't follow me :)

[Lex Savage @lexsavege]

@RightSideOfAI This has been a long-standing issue for DeepSeek. Since there is no specific version number in the system prompt and its data is outdated, it refers to itself as v3. I personally think they are secretly testing v4.1 and have quietly released it. Its reasoning process has changed.

[Lex Savage @lexsavege]

@teortaxesTex I've noticed that Deepseek's way of thinking has changed. It used to think like a normal person. But now it always thinks in bullet points. When it does a web search, it still thinks normally. I don't know if this is better but something is going on.

[Lex Savage @lexsavege]

@Elaina43114880 I've noticed that Deepseek's way of thinking has changed. It used to think like a normal person. But now it always thinks in bullet points. When it does a web search, it still thinks normally. I don't know if this is better but something is going on.

[Lex Savage @lexsavege]

@Elaina43114880 I hope so. I’m really curious. I want this more than anything 😅. My only wish is for the model to give smarter answers and be able to see images. And also web search in expert mode.

[Lex Savage @lexsavege]

@Elaina43114880 was pretty fast. expert mode did worse even though it thought more. in expert mode even though I didn't want another language it thought in english. in fast mode it thought in the language I asked. they are trying something in the background. I hope it has gotten better

[Lex Savage @lexsavege]

@Elaina43114880 Right now I said let me test the model on the deepseek site and these are the results I got. when I wrote in english the model did worse. here I wrote in my own language and I think it wasn't bad. of course it could be better but in fast mode the model thought for 23 seconds it-

[Lex Savage @lexsavege]

@teortaxesTex https://t.co/qyLcMXdsO2

[MiniMax (official) @MiniMax_AI]

MiniMax M3, Open-Weight, Now On Hugging Face , with only ~428B parameters and ~23B activated parameters

Weights:
https://t.co/g4Ybfa2kWH
MiniMax Sparse Attention:
https://t.co/HcTlWRotG3

[Lex Savage @lexsavege]

@scaling01 What if, a month later, some random Chinese model comes out with the same performance? 💀 A year is a very long time. I don't think it will take that long.

[Lex Savage @lexsavege]

@ErenChenAI I don't want to be misunderstood, I'm not a pervert or anything but I would give everything I have to own or even just touch a humanoid robot like that 😅. I hope I get to see such robots before I die. The Chinese are very good at this but they’ll definitely be expensive, though.

[Lex Savage @lexsavege]

@bridgemindai You shouldn't just look at benchmarks. Deepseek got better scores than Kimi AI, why aren't you looking at that? :D I think there's no difference at all. Judging based solely on benchmarks is pointless. Anyway, you can keep paying more for American models :D.

[Lex Savage @lexsavege]

@sheriyuo Grok-chan 😲 I asked him and he said that's how he sees himself. I did it with Grok Image. It seemed more appealing to me, but it turned out a bit exaggerated :d

[Lex Savage @lexsavege]

@Boxzmqook @Elaina43114880

[Lex Savage @lexsavege]

@Elaina43114880 Deepseek v4.1 flash is there?. how can you be sure?. I hope it is published quickly I would like it very much. Right now in the country I live in it is Wednesday morning 11 and today I haven't tried deepseek at all. yesterday I used fast mode a lot but is hallucinating a lot.

[Lex Savage @lexsavege]

@Elaina43114880 🤣

[Lex Savage @lexsavege]

@Elaina43114880 Me too 🙏. Honestly, DeepSeek is enough for me right now. The answers are pretty good and it's free. I can't try other American models because they're paid. I've tried other Chinese models but didn't like them much. Let's see what happens in the future.

[Lex Savage @lexsavege]

@Elaina43114880 I’m not Chinese but even I can see this. I think there’s at most a 1-2 month difference. The Chinese have really started to improve. I'm honestly rooting for the Chinese. At least they're doing it for free and open source. I can't bear to pay $30-200 for a single model 😅

[Lex Savage @lexsavege]

@Elaina43114880 I looked through all the articles online and they’re all comparing older models :D. For some reason they don’t include the DeepSeek v4, Qwen 3.7 series or Qwen 3.6 series. Instead they use the DeepSeek R1 or Qwen 3 models just to confuse people. +

[Lex Savage @lexsavege]

@michpokrass Unlimited free GPT 5.5 🙏

[Lex Savage @lexsavege]

@teortaxesTex Unrelated to the topic, but DeepSeek Expert mode can no longer perform web searches. And it can't view links 😭. I don't know why they're doing this. It was fine yesterday and the other days. I hope they fix it...

[Lex Savage @lexsavege]

@teortaxesTex They compared it with Deepseek 3.2. The V4 Pro is frankly better. No matter what anyone says, it's a pretty good model. I tried the Kimi K2.6 and GLM 5.1 models and they're bad. They're very slow. I can't use them :D.