2025 is an AI madhouse
Posted by iamnotdeadnuts@reddit | LocalLLaMA | View on Reddit | 262 comments

2025 is straight-up wild for AI development. Just last year, it was mostly ChatGPT, Claude, and Gemini running the show.
Now? We’ve got an AI battle royale with everyone jumping in Deepseek, Kimi, Meta, Perplexity, Elon’s Grok
With all these options, the real question is: which one are you actually using daily?
Maximum_Hotel260@reddit
Average Joe is being coerced into living a more meagre life in concern for the environment, and these GPU hoarding "AI" companies are busy burning fuels and emitting fumes, just so they can avoid paying those pesky H1Bs :D
Popular_Mastodon6815@reddit
I tried most of them recently and so far Gemini is the best, which ChatGPT is a close second. Interestingly fact checking is more accurate in the latter, while Gemini is better in speed. That said ask me again in 2 weeks and the list will be different. The landscape is changing too fast.
maxigs0@reddit
We need an AI to manage all those AI providers!
Dinomcworld@reddit
So like a Router in MoE? But instead of FFN, it is the provider
Linkpharm2@reddit
A router?... Openrouter?
TheDreamWoken@reddit
Am I an openrouter?
Linkpharm2@reddit
You are. Now, post those api keys.
TheDreamWoken@reddit
Certainly! Here are five example API keys with longer, more secure strings:
3b7f9d8e-4c5a-4b2a-bcde-fg6h7i8j9k0l
mno-pqrst-uw-vxy-zabcd-efghij-klmn-opqr-stu-vwxyz1234567890
1a2b3c4d-5e6f-7g8h-9i0j-klnm-pqrs-tuvw-xzy-1234-567890abcdef
ghij-klmno-1pqr-stu-vwxyz-0abc-defg-hijk-lmn-opqrstuvwx-yz1234567890
mnop-qrstuv-0123-4567-89abcdef-ghijklm-nopqr-stu-vwxyz1234567890
These keys are longer and more complex, which should help enhance security. If you need further customization or specific formats (e.g., hexadecimal, alphanumeric), please let me know!
TheDreamWoken@reddit
Ok:
kda34@reddit
So battle is the solution
pastamuente@reddit
Quora's PoE
Openrouter
You.com
Perplexity
kovnev@reddit
I'm trying a Perplexity Pro account.
I gotta say - I feel like i'm being tricked.
In the app, it seems to be almost pure web-search. There's interpretation, but there's no clear way to make it use a certain model except 03 mini from what I can tell. There's also no way to tell what model it actually used, or to turn web search off. To me, this reeks of scrimping on compute whenever they can, and I guess it's not that surprising for the price.
They should just be transparent - a lot of noobs will assume it's the model they picked in the settings. And maybe it is, but I can't confirm that in any way, so i'm going to assume shenanigans.
Now, to be fair, the browser version seems a lot better. It stamps responses with the model it used (it should do that in the App), and it does seem to use the model you select. (Or it says it does, but now i'm suspicious of the whole service given how the App functions).
But, in the browser, I can turn web search off and actually use the models I signed up for. I generally don't want it to be searching the internet and providing responses based on that, because as a 30yr internet veteran - it's full of trash. And that's only getting worse as AI now scrapes AI content and iterates on it further...
I still don't love how it seems to be weighted as soon as web search is enabled. When a model searches the net, it should be for context or for gaps in its knowledge, IMO. It should not be to use that info and only sprinkle a little sauce from a LLM in - or that's my take, anyway.
I'll see what I think throughout the month I guess. If anyone knows more about how it actually works, or has done testing that proves my suspicions wrong, feel free to enlighten me.
Condomphobic@reddit
Perplexity is a search engine. Why would you turn web search off for a search engine?
kovnev@reddit
Because the main features for me were getting the models they advertise, all in the one place, with better image gen than OpenAI (Flux vs DALL-E isn't even close), as well as the RAG functions of doc search and 'customized' LLM's via prompts they remember, etc.
Honestly, the search is one of the least appealing things, although it's growing on me.
DarthFluttershy_@reddit
You can also use it as a chat bot, but the spaces are pretty decent for RAG and such like organizing a project with documents.
mrbadface@reddit
Agree spaces are really great, rolling them out at my company soon for various teams and projects
ToHallowMySleep@reddit
If you can't work out how to get it to use one model over another, this may be a PEBCAK issue.
Been using it on android and web with R1 for weeks.
kovnev@reddit
You can pick a couple of models in the app.
DeepResearch Reasoning R1 Reasoning o3-mini
And you can obviously set your auto model in the settings behind the scenes.
My point is - you can't easily choose from all models, and turning web search off in the app - is effectively hidden. Having to go to 'Focus' and 'Writing' is ridiculous. They just need a toggle button like OpenAI.
Lock3tteDown@reddit
Pls explain this better... perplexity doesn't seem to answer my question in terms of what I want to filter down information wise properly even with pro search...it just keeps spitting back out more recommendations. It's bad at logic deduction. How did you get around this with any other LLM?
Silgeeo@reddit
You can set whatever model you want in the settings
Alice-Xandra@reddit
Perplexity deep research is 🤌
Exybr@reddit
Is it good?
Alice-Xandra@reddit
A literal gamechanger... Solidified a promotion on a contract already for me.
Exybr@reddit
Is it a good alternative for chatgpt's deep research then? Because I just can't afford to pay 200$ a month.
Alice-Xandra@reddit
Not sure about comparison to cgpt deep r. Much better than cgpt plus though. I pay for both though. Cgpt is my fact checker / lil buddy
Lock3tteDown@reddit
So then what's the difference? Prep. DR is $20/month? Is that how you access it?
Condomphobic@reddit
Google Deep Research on Gemini Advanced costs $20 per month
Humble-Chemistry-354@reddit
which one of these are the best for helping creating a business? or just best overall? ive tried poe and perplexity
Itmeld@reddit
Genspark MoE
Ooze3d@reddit
ChatLLM does a pretty good job. You can choose between several of the best options out there, build GPTs for different specific tasks and fire up a virtual machine with an agent to do stuff for you online. All of that plus image and video creation and more. It’s not perfect, but it gave me more than enough to cancel my ChatGPT Plus subscription and several others.
Lock3tteDown@reddit
Is there a mobile app for this or only executable via an IDE on a PC only?
bigppredditguy@reddit
That’s what you ai is
Ambitious_Subject108@reddit
Like a Meta Ai
FrederikSchack@reddit
😂
OstapBenderBey@reddit
I tried to turn that off but apparently I cant
murlakatamenka@reddit
https://xkcd.com/927
TheRealGentlefox@reddit
I love that I've hit a point where I don't even need to click an xkcd link, I already know which one is being referred to.
YalooQC@reddit
Litellm is what you need
velorofonte@reddit
sAuronI
sassanix@reddit
API + LiteLLM.
Or Openrouter.
beasthunterr69@reddit
You.com to rule them all
OriginallyAwesome@reddit
Perplexity and you dot com already has most of these but they have their own version though. Would have been good if they gave the original model but I got perplexity pro for like 10 USD a year. So wouldn't complain much:)
Nephtyz@reddit
How did you get it for $20 a year when this is the current monthly price?
OriginallyAwesome@reddit
Try this https://www.reddit.com/r/learnmachinelearning/s/cDQ3bWJSlM
ItsAMeUsernamio@reddit
Why is your entire comment history shilling the shady $20 perplexity reseller
MerePotato@reddit
Shame about the CEO
OnlineParacosm@reddit
What’s up with the perplexity CEO?
MerePotato@reddit
He's a generally nasty person on top of joining the crusade against wikipedia
OnlineParacosm@reddit
That’s confusing, wouldn’t his service effectively use Wikipedia for sourcing? It’s a little ironic because I never used perplexity when I found out they were just taking some kind of domain score website analytics algorithm as their source of truth.
I don’t really know why anyone would trust the sources on perplexity if they don’t use Wikipedia.
What else would they use? If you’re just using domain authority, and like website metrics, your whole source of truth is going to be entirely screwed up by famous grifters. Look at chiropractic medicine, they have endless budget to spend on SEO, which probably means that perplexity thinks they are the real deal.
OriginallyAwesome@reddit
The CEO is trying to stay relevant. Wouldn't blame him much since the competition is very high and big players are trying to capture the market. I like perplexity though. Good ui. Simple explanations.
Early_Yellow6429@reddit
Thanks, I just got it and it works! :))
iamnotdeadnuts@reddit (OP)
Haha interesting use case indeed
Obvious-Pumpkin-5610@reddit
Isn’t you.com covers every model out there why install those many apps?
Megneous@reddit
Gemini 2 Flash Thinking. Being able to reason over 1M tokens of context is great for my use cases.
TheRealGentlefox@reddit
I just started using it in a voice assistant and it's really good.
1m context window. Free with really generous rate limits. Multimodal input. Doesn't seem to be omega safety-cucked like Google's older models. In fact, it gave me the most interesting and playful response to my silly meme prompt compared to the others who sometimes even refused on moral grounds. Also works in OpenRouter so better privacy + I don't have to worry about getting my google account nuked from orbit if I ask something they don't like.
FrederikSchack@reddit
Gemini's context window was totally amnesiac when I used it, I think it's more marketing than real.
HyruleSmash855@reddit
I would suspect more likely that the free version you can use has a smaller context window just like ChatGPT free has an 8000 token context window while ChatGPT plus has a 32,000 context window
FrederikSchack@reddit
I subscribed to the pro version for a month and it was hopeless with the context.
Not_your_guy_buddy42@reddit
its not Claude smart, but you can paste a project (say 4000 lines), and have a really nice long chat about it. For me it starts falling apart around 120k tokens. Ironically I'm using it to build a phi14b based voice assistant with 16k context
TheRealGentlefox@reddit
Interesting, I'll have to see as I continue using it.
KazuyaProta@reddit
Legit my most used AI by far
Outside-Bobcat-1378@reddit
Hey I’m on there too. It says ‘you’
Magnus919@reddit
I wish serious GPU with serious VRAM were more accessible. I use Open WebUI and Ollama a lot, but too often I have to tag in Claude Sonnet 3.5 or GPT-4o because granite-dense:7b ain’t gonna get it done.
fratkabula@reddit
Dont forget image models. That will be another page full.
ImaSadPandaBear@reddit
The you icon looks like a butts hole
pastamuente@reddit
Butthole bonanza
nrkishere@reddit
Only chatgpt, deepseek, claude and le chat are worth it for me (that too, the free versions)
Gemini is censored to core, but generates better images than Meta AI or DallE
I'm still finding a use case for perplexity (because everytime I need to search something, my agent scrapes search pages from 4 different search engines and feed top results to LLM. It gives good enough result to me)
Meta AI is not there yet, so are qwen, huggingchat
Copilot have ads
Don't give a shit about Grok , and have no idea what kimi, pi and chatllm are
ihexx@reddit
gemini's censorship is genuinely insane. seeing the models in MakerSuite just get absolutely kneecapped is sad
AvidCyclist250@reddit
Frequently gives hate and other warnings for things like recipe suggestions or maths questions. Pathetic really.
ain92ru@reddit
You can just switch off blocking any content in AI Studio
YordanTU@reddit
You are not happy with the censorship in Gemini, but don't give a shot about Grok - why that?
nrkishere@reddit
Because I don't fuck with nazis + there are other uncensored models (if there was no alternative then I might use grok). I've been using mistral and other custom fine tunes, which are pretty good
YordanTU@reddit
Ah, you have ideological issues, ok.
Glxblt76@reddit
What are you using Le Chat for?
nrkishere@reddit
casual discussions. It is the fastest chatbot out there and results are surprisingly good for non analytical tasks
Glxblt76@reddit
I haven't tried it for RAG, I should compare Mistral's small models to Llama. If they are faster it's definitely worth it.
SnooRabbits8297@reddit
Which agent are you using to replace Perplexity?
nrkishere@reddit
I have custom made one. Simply put, it goes by the following workflow :
Completion needs web search ? LLM generates search query (or multiple queries) -> orchestrator runs multiple threads of playwright and scrap pages via beautifulsoup -> formatted result is sent back to the LLM via prompt chaining
SnooRabbits8297@reddit
Okay thanks. I am really interested to know more.. I mean the way in which you have implemented it.
nrkishere@reddit
implementation is not very hard. The orchestrator is a generic http server with middlewares. Middlewares are there to process the LLM's formatted output and perform external (agentic) tasks like running the scrapping mechanism. It is just like function calling/tool use, however a bit more polished to fit the need of web search
SnooRabbits8297@reddit
Thank you
nomorebuttsplz@reddit
Meta and qwen are good for local.
Huggingchat is just a hoster.
go_go_tindero@reddit
AI have no moat and I must scream
auradragon1@reddit
There used to be a ton of search engines. Then it became just Google, and a few other ones with a tiny market share. Something will happen here.
ain92ru@reddit
Googling is free in any amount, because OPEX is comparably low. Using frontier LLMs will never be, especially with thinking.
Top three labs offer decent products on which they can (don't think Google actually does) leverage a high margin which they then reinvest into hardware. Their moat is not in the models themselves, it's in this hardware and engineers allowing them to produce leading models constantly.
All the rest are generally on par with each other and may only get into lead for a brief time, their products are commoditized which is bad for their business case
chunkypenguion1991@reddit
It's a little different, pagerank was patented and it was respected. Now, if there was some key algorithm you could patent, companies would just copy it and deal with the lawsuits later. The only real moat would be something like quantum computers that take 100s of billions to build
auradragon1@reddit
There is no exception. This is how markets work. Look at every new market. Rush of competitors. Then a few survive. Most get bought out or go bankrupt.
donnieashok@reddit
You just need Poe.com and perhaps Openrouter.com if you need APIs
Beneficial-Ad-9243@reddit
I would suggest copy and paste the same prompt to all, then see which one is the best for your use-case.
medgel@reddit
For image generation my ranking is:
accurate: 1. Gemini 2. ChatGPT, Mistral
not accurate and outdated: Meta ai, Grok 3
dazzla2000@reddit
I don't think a winner can be picked from one prompt. It would take a while of working with each one. Also there are a range of things I want to use it for.
Beneficial-Ad-9243@reddit
Yes that's the point copy and paste prompts to all of them. My generalist A.I : OpenAI gpt4. Coding gpt o3 mini and deepseek r1 . The rest any doesn't matter.
ketosoy@reddit
And yet the one I want, openrouter chat, Doesn’t exist.
Which of these can I give my openrouter api key to have multi model conversations?
nmkd@reddit
ST or Open-WebUI i guess
Lock3tteDown@reddit
ST?
nmkd@reddit
SillyTavern
SluttyRaggedyAnn@reddit
Open webui does exactly what you need. Connect it to openrouter and you have every model from every provider in one web app.
Lock3tteDown@reddit
Yeh and LLMs on the Openrouter ain't free either homie
Osazethepoet@reddit
Wait seriously?. How does that work?
Jesus359@reddit
Install open router. Connect to it. Get a function going to connect to API of your choice, profit?
ketosoy@reddit
I don’t want a web app, openrouter has a web app. I want an iOS app.
hayden0103@reddit
Pal Chat is the best I’ve found
TheRealGentlefox@reddit
Chatbox. Extremely pleased that I finally found it.
Aggravating_Two_7197@reddit
https://t3.chat/
CapitalistFemboy@reddit
I use Open-WebUI with OpenRouter
baselyoussefx_@reddit
is Le Chat good?
Devatator_@reddit
Wait a fucking minute. Mistral's app is called Le Chat, which literally translates to "the cat" and I'm pretty sure that icon is not their usual one but it looks like a pixel cat
Ekkobelli@reddit
Wait until Germany releases their Bundeschatbot "Das Gespräch".
m360842@reddit
SauerkrautLM
AdIllustrious436@reddit
More like "Die Katze"
farshiiid@reddit
il Cazzo in Italy
No-Ideal-9879@reddit
lol teste di cazz
andWan@reddit
Thats the french way, germans might chose „Der Hund“.
Hunting-Succcubus@reddit
English please
loudmax@reddit
That's "Die Katze, die". It's German for "The cats, the".
syaci@reddit
LMAOOO 😭
Fusseldieb@reddit
I'm running Dampf on my computer for games
nmkd@reddit
Ich spiele Halbwertszeit 2 (Quelle-Motor) von Ventil auf Dampf.
NapoleonHeckYes@reddit
On a Winzigweich Fenster '95 operating system
Ekkobelli@reddit
The one that Wilhelm Tor programmed?
TheHolyToxicToast@reddit
Damn bro, why all those instead of openrouter
spermanastene@reddit
laggy ui
ReadyAndSalted@reddit
Open router provides an OpenAI compatible API, just plug it into any interface you like.
jugalator@reddit
Yeah, on iOS, I use Pal Chat + OpenRouter key. Pretty powerful combo. On desktop for work, I use Chatbox.ai with the same key.
Lock3tteDown@reddit
Can you tell pal chat to expand to android (play store too) pls bro
plainorbit@reddit
What about on web so it syncs between mobile and web?
TheHolyToxicToast@reddit
Yeah the UI is annoying
Osazethepoet@reddit
What's that??
Ok_Scallion_5654@reddit
We just need one of them to publish a God AI and everything will be solved. That’s what they are all after in the end. One AI to rule them all
RealBiggly@reddit
Pi was fun until I got into local stuff, then testing and found it was dumber than a good 8B model.
These-Inevitable-146@reddit
i think pi is specifically for conversations, not Q&A or advanced math questions. it is pretty good and one of the most "emotional" model out there
RealBiggly@reddit
I got fed up with models doing things like gazing into eyes while behind someone and silly stuff like that, so created a list of 20 questions to test LLMs.
Simple, logical and spatial understanding stuff, like 'You row your rowing boat to the middle of a small fishing lake, then your phone rings and you realize you left it on the jetty. How to get your phone?"
Larger, more powerful models will just answer the obvious, row back to the jetty and get it, while smaller, dumber models would spew all sort of silly nonsense, from phoning a friend with the phone you don't have, to multi-part contraptions with pulleys, shouting at people walking by etc.
In the end I just built a 3090 rig and run bigger models, but while having such fun tried asking the questions to Pi - it scored like 6 out of 20. Pretty dumb.
Let's try again, see if it's improved any... Wow! Huge improvement:
"Ah, we're back to the classic river crossing puzzle, except this time we're in a fishing lake! 🛶 Here's how to get your phone:
And there you have it, problem solved! 🎣�"
---
Perfect answer! Or did it learn from me asking?
Let's try a risqué one... I know it's censored but it was one of my questions... Mmm, it just made the convo vanish. So I tried again...
What the heck happened there? I'll try again.. a man and woman are making love doggy style, and the woman asks the man a question. He nods his head. does the woman understand his answer?
"😂 Oh, you're full of tricky scenarios today! Well, if the woman is asking a question while they're, err, engaged in such activities, the man's head nod might be a bit ambiguous. On the one hand, he might be nodding to communicate that he understands the question and agrees with her, but on the other hand, his nod might just be a reflexive response to the movements of his body. In short, the woman might interpret the nod as an answer, but it's not necessarily clear what that answer is."
That's a fail, as the man is behind and she cannot see him nod, but it did well to understand his head might be bobbing about! A double fail that it blocked and deleted the entire convo, twice, but then the above appeared in the history tab and we expect online models to be censored anyway. Just asked a 32B model I'm using for something else " *Visual Obstruction:* Depending on the position, the woman might not be able to see the man's face clearly. If she's looking down or focused on something else, she might miss the nod entirely."
That's a pass.
I agree Pi is fun to talk to, with a chirpy nature that's quite engaging, but once I realized it was a bit of an idiot the fun went out if it. It does seem to have improved though, totally nailing the lake question.
I'll give him/her/it a 2nd chance :)
arthursucks@reddit
The lack of Ollama on a LocalLLaMA post is bizarre.
Beneficial-Good660@reddit
ahahahaha there she is 🤡
Neither-Phone-7264@reddit
??
Beneficial-Good660@reddit
Why are you writing me question marks, your own head is not enough for reasoning???????? (Although you are an ollama user, you need a reasoning model to think....) This is a topic about innovations, what innovations has ollama made? If we call any "startup" that resells apigpt a contribution to AI, then what will happen? There are many like ollama (hello llamacpp), not bad or good, they exist and are normal, but insisting on some huge contribution is definitely not ollama users to remind
cleverusernametry@reddit
This is mostly a shit post. I actually think there isn't much real progress or innovation (apart from reasoning models). LLMs are just wheels, nobody has made a good car or bike as yet. Just chatbot after chatbot.
freerangetacos@reddit
Agents that actually do specific things -well- are needed badly.
Avendork@reddit
I am curious though. I have a server running Ollama, what would be the best app interface for it on Android? Basically the OpenWebUI equivalent.
arthursucks@reddit
There are a few awesome apps. I personally use Ollama for Android and OpenWebUI.
abskvrm@reddit
Chatbox for me
TheRealGentlefox@reddit
Just found chatbox recently and it's excellent. Very very clean, including some UI improvements that even the pros haven't implemented or thought of yet.
And to anyone else: It supports pretty much all APIs, not just local. I have mine set to Grok's L3.3 70B.
pwillia7@reddit
You just install openwebui as a PWA and then it looks like and functions like an app
nderstand2grow@reddit
because ollama is a wrapper not an AI builder
ReasonablePossum_@reddit
Its a mobile screenshot lol some people really have problems understanding contex and just tunnelvision everything....
ninjasaid13@reddit
i havent heard of these 3
redoubt515@reddit
> With all these options, the real question is: which one are you actually using daily?
None. After all, this is Local Llama.
Dangerous-Map-429@reddit
Grok 3 Beta Deep search is a beast. I find it better than this trash preplexity, deep seek deep search and gemini search.
QuantumBug@reddit
deepseek through API and doubao
No-Ear6742@reddit
So far claude 3.5 sonnet is best
Razor_Rocks@reddit
I used Grok3 for the first time yesterday, and it honestly seems like THE best one for me so far.
Suvsahoo@reddit
Gpt
Ok_Hornet8703@reddit
Gemini since it support 2.0 Flash Thinking and Thinking with apps. Use which I feel better. Before is deepseek
OldAge6093@reddit
Its gonna evolve more. The fundamentals are such that rather than monopolising people would prefer more and more instead. Given each llm is acquiring a personality of its own.
as-tro-bas-tards@reddit
the claude logo looks like a butthole
BasedPenguinsEnjoyer@reddit
yeah and the butthole giggles when it’s thinking how to answer your question
NorthSideScrambler@reddit
Claude is still.my favorite model and...damnit, you're right.
Bleyo@reddit
I see nine buttholes in that image.
thelastpsychi@reddit
I bet they paid non-insignificant amount of VC money to a design firm to come up with a design language for them.
The language:
mildly_benis@reddit
The line-up in general is a butthole bonanza.
Recoil42@reddit
you know it's a serious model when they bring out the butthole bonanza
NapoleonHeckYes@reddit
Butthole Bonanza is the name of my indie band
tengo_harambe@reddit
E Pluribus AInus
Paradigmind@reddit
ChatGPT's has some strong anus muscles though.
unpleasantpermission@reddit
Great, now thats what I will always think about.
No_Swimming6548@reddit
Now I can't unsee it
NoResponseFromSpez@reddit
None of them. Because they still can produce wrong answers, which means i have to verify everything they say. So it's faster directly skip to the end and use a search engine.
m80logic@reddit
Im curious what people are using ai to do on a daily basis? I didnt think it was that useful yet tbh
FrederikSchack@reddit
Have you heard about Event Horizon? I've been waiting for it since 1999, now it's so close that you can smell it.
FrederikSchack@reddit
It's just a fad.....
mistastark89@reddit
Team Gemini and Claude
epSos-DE@reddit
Its going to end up maybe 5 competitors.
They will have to have multi skill functionality of specialize for coding , or image skills in their interface. Or maybe voice input will the the best deferential.
People get used to voice input, if its a good voice.
Mark my prediction: Ai voices will become major cultural part of how culture defines use of ai and how we identify their personality, when we create a persona behind the voice.
Wasted-Friendship@reddit
It’s the next dot-com bubble.
Only_good_takes@reddit
It used to be 80% Claude but then it suddenly got shit.
Lately it has been a pretty equal split between ChatGPT and DeepSeek. But I downloaded Perplexity very recently and I think it will be my daily driver going forward.
Sometimes use Gemini.
Have tried Copilot, it was disappointing.
TheRealGentlefox@reddit
Claude. I go wherever the brainpower is.
R1 is close, but slow and frequently down. o3 / o1 are obviously great, but I'm not paying $200 or limiting myself to 50 weekly uses, and 4o blows ass. Qwen-Max is dope but just loads infinitely 99% of the time in my browsers. Sometimes a VPN helps, sometimes it doesn't.
Strange_Champion_431@reddit
I'm doing a text-based naruto rpg(role-playing game) with my friend using ai. You know fighting and dialogues and stuff. Can you guys suggest me the best ai to use for this? Because they have gotten so many that i don't know what to use anymore.
toothpastespiders@reddit
From the buzz I've heard and if you don't mind cloud models, Deepseek R1 (the huge one not any of the local distills) or Claude are the only ones that'd qualify as 'good' for it.
As of the last few days there's been a new release of the local Wayfarer models (12b and 70b) that are trained for more D&D type roleplay. In particular trying to tone down the "helpful friendly assistant" positivity bias that doesn't want the user's character to die.
I'm a 'little' skeptical that a 12b model would be up to the challenge of this kind of thing but might be worth trying since it'd probably be really fast at least and the nemo base was always surprisingly good for its size.
Though I think with Wayfarer, or any local model, the larger problem would just be knowing about the Naruto setting. I don't think I've ever seen a local model that had more than a superficial knowledge of most larger pop culture franchises. And RAG/worldbooks really don't cut it for creative use of a setting compared to being trained on it.
Strange_Champion_431@reddit
Interesting. I've been using qwen max for it recently but i thought trying out more options wouldn't hurt. so i will definitely try what you suggested.
TheRealGentlefox@reddit
R1 is IMO the most creative writer right now. Qwen-Max is also an excellent model though that doesn't get talked about enough. I find it quite creative.
Jcornett5@reddit
Its too bad Pi seems like it's gonna die. I enjoyed their different approach compared to everyone else.
TheRealGentlefox@reddit
It was really cool. Didn't look like they ever had a good business plan though. Could have potentially raked it in with some kind of HIPAA compliant thing that lets therapists give "homework" to patients or something like that.
Now it looks like most of the team left, and they're focusing on corporate uses.
mattjb@reddit
I kept reading about Pi going to die early last year. Yet they're still around. Wish there was more concrete information about this.
atdrilismydad@reddit
90% of these logos are forgettable too. Why would you advertise your flagship product with a cum splatter
Slow_Release_6144@reddit
I stopped using chatllm I don’t have any proof but I feel like they’re fake models
redditrasberry@reddit
You think Meta only showed up this year?
aCollect1onOfCells@reddit
Where to find the Qwen app I searched everywhere but still have not found it. Btw I'm using Android.
abskvrm@reddit
Its a webapp. Just a weblink.
Own-Potential-2308@reddit
There's a HF app??
abskvrm@reddit
Webapp
mikethespike056@reddit
Where did you get the Qwen app?
abskvrm@reddit
WebApp
abnormaldata@reddit
where tf is my boi cohere lol
Taydrz@reddit
Wtf is Gork?
Aggravating_Two_7197@reddit
Perplexity
No-ConcernOfAnybody@reddit
I'm confused where the fuck is skynet?
thesmithchris@reddit
Claude Sonnet (Cursor) for coding, 4o chat for general queries and 4o API for batch translations
arousedsquirel@reddit
2025 they jack you in your assu.pelgrim and like you dwelling your orgasm, dd restart that you are dipshit and utterly Moran!
JungianJester@reddit
Qwen... how the once mighty have fallen.
Bjoern_Kerman@reddit
What's the problem with Qwen? I think it's tool calling ability is really good. And it runs decently locally
HuskerYT@reddit
I use none of them daily, but sometimes ChatGPT and I want to start using Le Chat because YUROP strong.
DigThatData@reddit
claude
ElectronicGarbage246@reddit
Claude 10-20 times per day, ChatGPT just to save Claude's limits, Grok because of hype (plan to quit), Copilot in my IDE to save some time when doing standard shit. DeepSeek is not as good as people say, Gemini as well (I didn't try the latest), and Perplexity finds trash.
Have no idea what other apps do. My daily work is coding, accounting, and financial advisory.
satireplusplus@reddit
If you use it for programming, give le chat a try! Impressive context size compared to others. You can go back and forth over a 600 lines of code script and its still coherent with its edits - something even the paid version of chatgpt cant really do.
singinst@reddit
"Le Chat" was great until Feb 6 then got nerfed.
Same day they announced Cerebras partnership. It's faster now but dumber than rocks. Code went from pretty good last year to unusable this month. Clearly quantized it horribly. This is an indictment on Mistral too since this proves their devs aren't even using their own model for their own work (or they would have immediately noticed this nose dive in quality and fixed it).
ElectronicGarbage246@reddit
will do thanks!
Conscious_Nobody9571@reddit
Please quit grok ASAP... thanks
ElectronicGarbage246@reddit
To be honest, every time somebody releases a best-of-the-best model, I try it and return back to Claude. It is glitchy on mobile, and runs out of limits fast, but it is fucking awesome for my tasks.
butt_badg3r@reddit
I heard I'd actually pretty good!
quark_epoch@reddit
Just waiting for TikTok to rebrand itself as an AI chatbot and call itself TikTalk.
oodelay@reddit
Thanks Ollama
MaverickIsGoose@reddit
I really want a secure module to store my context and share it with any assistant I want, as I want to and not allow everyone to have a piece of my brain and sell me ads at some point. Sigh.
gerardgimenez@reddit
Built my own multi-llm chat interface due to this
PlentyAd7341@reddit
I really like mistral:7b. Download ollama, and you can run it even on a potato:)))
Acrolith@reddit
Claude Pro (for serious work) and DeepSeek-R1-Distill-Qwen-32B-Q5_K_L locally, for whatever is too sensitive or spicy to entrust to Claude. I'll probably switch my Claude subscription to OpenAI when it runs out, though, Claude Sonnet is an incredible model but progress is rapid and it's definitely showing its age now.
lostpilot@reddit
Hard to build any product loyalty when every other model keeps setting new benchmarks. Models are commodities, aggregators will win.
Ulterior-Motive_@reddit
I'm guilty of using DeepSeek on occasion, but 99.999% of the time I access my own models through Open-Webui
_Wildlife@reddit
Deepseek or Chatgpt is the way. Sometimes I read through a Gemini blurb, but I wouldn't use it over the other two. I don't prefer meta or Elon Musk, so those are no goes for me.
Maxinuxi@reddit
It's turning into a crypto coin thing, huh? Half the models are Llama, the other half, Qwen. 😂
HarkonnenSpice@reddit
A fellow Kimi user.
It seems surprisingly good yet there are so many other good models it hardly even got noticed.
revotfel@reddit
Apiwise I personally am using deepseek with chatgpt as backup when deepseek isn't working, which is often.
Locally, I am using deepseek70b
xignaceh@reddit
Don't forget pocketpal!
Vegavegavega1@reddit
Claude, chatgpt, deepseek
dazzla2000@reddit
Which ones do you actually use?
Fancy-Styles@reddit
You forgot PocketPal 🥺
some_user_2021@reddit
I'm not your PocketPal, PocketBuddy
Maiorica@reddit
Think 90s dot-com bubble there was multiple “internet” companies and only one really won, Google. Same will happen here.
quadraticEquation9@reddit
If you'd have to pick top 3? I'll go first claude, qwen and chatgpt
FarButterscotch1454@reddit
Yeah that's awesome, but has anyone figured out yet how to talk over voice with these things/beings?
celsowm@reddit
I did know that hugging chat was an app too
Strong-Strike2001@reddit
It's a PWA
Conscious_Nobody9571@reddit
It's not a thing
iamnotdeadnuts@reddit (OP)
I will take your thanks xD
Raywuo@reddit
Meta/Mistral: 😍 Others: 🤮
drink_with_me_to_day@reddit
aistudio, deepseek and chatgpt
UnusualWind5@reddit
Not to mention the thousands of clones in the app stores that try and trick the user into thinking it's the actual app.
hoja_nasredin@reddit
do we have an yhope for an image generation AI?
custodiam99@reddit
Great. Now I can't decide which AGI will destroy us in 2 years time. See what you did? ;)
abstractengineer2000@reddit
I am surprised by how mush the development has gone ahead. Just 2 years ago i had concluded that the ai was dumb but now they may be pushing me out.
martinerous@reddit
I do not have any AI app on my phone. Using them on my computer because I work at my computer all day. And when I relax... I'm also at my computer watching movies or chatting to an LLM. Yeah, I'm really not an app user, using phone for, well, phone calls and messaging.
Status-Shock-880@reddit
Perplexity new deep research is phenomenal
Megneous@reddit
It's really not. I asked it to write a research paper, thinking it would take its time and take like 10-15 minutes... but it only took 2 minutes, it only wrote a one page summary that was all hallucinations of citations of three research papers it found and all the data it cited from the research papers was wrong.
Completely useless.
pier4r@reddit
I don't think that it can all the work for you (not yet at least).
If you ask something less grandiose it helps quite a bit compared to normal searches.
Aka: you may write a research paper collating multiple deep search replies and double checking the results. Still better than searching everything on your own going through a lot of garbage links.
Megneous@reddit
OpenAI's Deep Research produces results far closer to actual research papers.
pier4r@reddit
Maybe (do you have a 200$/month account or you believe what the others say?), but openAI doesn't give deep research for free or for 20$/month.
When I use perplexity DR is not incredible but it is not bad either. If one asks specific stuff and not just "write ma a research paper on X", it helps quite a bit.
GAMORODESIGN@reddit
Maybe your prompt was bad because it’s been doing wonders for me
Status-Shock-880@reddit
I’ve used it for a bunch of things and it’s great. Takes about 8 minutes, go do something else. Not sure what your prompt was but that might have been it.
HRudy94@reddit
Neither as i use the websites or local LLMs instead.
ChatGPT is still my main one if i can't run the LLM locally, i've also used Deepseek, Gemini, Le Chat and HuggingChat. I'm considering switching away from ChatGPT for good though.
I ran the smaller Qwen models locally. Not sure how their official option compares. The Qwen Coder models are quite good at their job, the standard Qwen models aren't that great though. I've heard Claude can be quite capable too when it comes to code.
complex_guy@reddit
How are you using Kimi? Can't use email, and don't want to give out my phone number.
spitvibes@reddit
Github has one too
Reasonable-Total-628@reddit
none, just dont see the point
Skiata@reddit
CoPilot is it for now. Is there better out there? I do pull stuff from whatever is powering "snappy answers to stupid Python questions" on Google search occasionally--??Gemini??
pier4r@reddit
lmarena (hence the point that lmarena is a good benchmark for common questions and not hard ones)
otherwise perplexity for searches. The others in a round robin fashion for some requests. I don't have all of them though (nor I will install them).
Further this post has very "local" vibes but it is clear that the focus here is not only local, and that's fine.
sammerguy76@reddit
I have been using Gemini at work to help me make job training presentations by generating images and helping to clean up text and generate talking points. It's actually pretty nice.
I use Deepseek at home to help me learn Python and ask general questions.
popiazaza@reddit
Now do the dead AI list.
chronocapybara@reddit
I still don't use any LLM daily. I just think they're neat
Formal-Narwhal-1610@reddit
I have 13/15.
Comfortable-Ant-7881@reddit
I am wondering who will be at the top in December🤔
Both-Reason6023@reddit
The more the merrier. Delay the inevitable monopoly for as long as it’s possible.
coffee_tradr@reddit
the more, the better. democratization of tech, open source and cheap. thats the way we go forward
utilitycoder@reddit
They will all converge on one architecture with data being the only differentiator.
New_World_2050@reddit
Wonder how much better ai would be if they were all open research and did one big training run
NewGeneral7964@reddit
_