/lmg/ - a general dedicated to the discussion and development of local language models.

Previous threads: >>106217978 & >>106212937

►News
>(08/06) GLM-4.5V released, based on GLM-4.5-Air: https://hf.co/zai-org/GLM-4.5V
>(08/06) Qwen3-4B-Thinking-2507 released: https://hf.co/Qwen/Qwen3-4B-Thinking-2507
>(08/06) Koboldcpp v1.97 released with GLM 4.5 support: https://github.com/LostRuins/koboldcpp/releases/tag/v1.97
>(08/06) dots.vlm1 VLM based on DeepSeek V3: https://hf.co/rednote-hilab/dots.vlm1.inst
>(08/05) OpenAI releases gpt-oss-120b & gpt-oss-20b: https://openai.com/index/introducing-gpt-oss

►News Archive: https://rentry.org/lmg-news-archive
►Glossary: https://rentry.org/lmg-glossary
►Links: https://rentry.org/LocalModelsLinks
►Official /lmg/ card: https://files.catbox.moe/cbclyf.png

►Getting Started
https://rentry.org/lmg-lazy-getting-started-guide
https://rentry.org/lmg-build-guides
https://rentry.org/IsolatedLinuxWebService
https://rentry.org/recommended-models
https://rentry.org/tldrhowtoquant
https://rentry.org/samplers

►Further Learning
https://rentry.org/machine-learning-roadmap
https://rentry.org/llm-training
https://rentry.org/LocalModelsPapers

►Benchmarks
LiveBench: https://livebench.ai
Programming: https://livecodebench.github.io/leaderboard.html
Code Editing: https://aider.chat/docs/leaderboards
Context Length: https://github.com/adobe-research/NoLiMa
Censorbench: https://codeberg.org/jts2323/censorbench
GPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference

►Tools
Alpha Calculator: https://desmos.com/calculator/ffngla98yc
GGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-Calculator
Sampler Visualizer: https://artefact2.github.io/llm-sampling

►Text Gen. UI, Inference Engines
https://github.com/lmg-anon/mikupad
https://github.com/oobabooga/text-generation-webui
https://github.com/LostRuins/koboldcpp
https://github.com/ggerganov/llama.cpp
https://github.com/theroyallab/tabbyAPI
https://github.com/vllm-project/vllm

►Recent Highlights from the Previous Thread: >>106217978

--VLLM CPU offload fails due to missing pre-built support and poor error feedback:
>106218692 >106218723 >106218737 >106218783 >106218827 >106218918 >106219093 >106219192
--Qwen 235b sampler tuning for coherence and variety:
>106221308 >106221341 >106221549 >106221619 >106223410
--RAM increase negligible for llama.cpp when still relying on disk swap:
>106222670 >106222743 >106222860 >106222753
--AI solves Hitler riddle with safety filters interfering despite correct logic:
>106218100 >106223804 >106223902 >106224022 >106223851
--Hidden character card fields in SillyTavern causing unexpected prompt behavior:
>106219739 >106219801 >106219849 >106219871
--Merged PR simplifies Mistral model integration in llama.cpp by removing redundant runtime deps:
>106222328 >106222396
--GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models:
>106223114 >106223640 >106223691 >106223748 >106223785 >106223809 >106223811 >106223826 >106223212
--GLM-4.5 Vision performance and local model competition in roleplay and OCR tasks:
>106222085 >106222306 >106222378 >106222397 >106222829 >106222884 >106222909 >106222926 >106222738 >106222886 >106222920
--DDR5 bandwidth gains limited by NUMA and PCIe bottlenecks in MoE model inference:
>106221496 >106221688 >106222050
--Combating repetitive text generation with mirostat and context variation techniques:
>106221963 >106221992 >106222007 >106222019 >106222025 >106222039 >106222133 >106222149 >106222164 >106222177 >106222217 >106222232 >106221994 >106222003 >106222056 >106222059
--Mitigating repetitive speech patterns in long-running AI roleplay sessions:
>106218114 >106218192 >106218232 >106218261 >106218326 >106218526
--AI hardware bandwidth limits and co-packaged optics explained:
>106218178
--Miku (free space):
>106221496 >106223575

►Recent Highlight Posts from the Previous Thread: >>106217984

Why?: 9 reply limit >>102478518
Fix: https://rentry.org/lmg-recap-script

Returning newfag here. What's the meta 4090 model for RP/storytelling these days?

>>106225483
GLM-4.5-Air

>>106225483
OpenAI's GPT-oss 20B.

>>106225483
check this page to find out: https://www.reddit.com/user/TheLocalDrummer/submitted/

'toss in the 'rash

>>106225556
GOTEM

>>106225556
How will moatboy recover from this one?

>>106225495
>>106225500
>>106225524
Thanks anons!

Mikulove

>>106225627

>>106225627
hi butifel show bob

>August 11, 2025
https://archive.org/details/meta_leaks_part_1

>>106225655
>not behemoth/llama4.1
I am disappointed.

>>106225655

>>106225641
This time it is the local fauna harassing the Miku

How would you measure the amount of variety between rerolls for a given model+sampler?

Here's what I'm considering:
>have the model generate a story
>cut off the story partway through and have it generate N different completions
>give an llm judge the common prefix and two completions and ask it to rate whether the second completion plagiarized the ideas of the first
The goal is to detect how often rerolling just gives you the same thing in different words.

I know LLM judges are not very reliable, so I'm also wondering if it would be possible to do this by looking at embeddings or probability distributions somehow. For example, from the common prefix, complete N paragraphs + "[Character] knew he should ___" and see the distribution of things the model thinks the character should be doing at this point. Rerolls that are wildly different should have different distributions of what happens next.

>>106225647
>>106225703

>>106225727
I'd do something like generating multiple completions, sentence embedding each, then look at distances in the embedding space (TSNE, PCA, etc. could be used here).

>still no github issue for Air Vision
ACK

>>106225655
archive org is slow ahh https://files.catbox.moe/tlejyd.pdf

My hype list, from most interesting to least:
>Anthropic leak
>DS V4/R2
>Some new big Chinese model out of nowhere
>Grok2/3
>Whatever cohere is making
>Gemma
>Llama

>>106225876
Mistral Large 3 is off the charts

>AI sees quantum leap with GPT-3 generation, effectively destroying the Turing test and promising to usher in technological utopia.
>AI companies fall for the pajeet meme in their haste to rapidly expand to meet demand.
>3 years of endless benchmaxxing and safetyslopping later and it hits a wall.
Sadly the fate of every industry it seems.

>>106225921
AI doesn't need big jumps. This is fine.

>>106225921
I'll give it some credit, at least it demonstrated some degree of usefulness before hitting that wall
The current state of LLMs won't outright revolutionize things, but they'll have a place until the next major development comes along, whether that's still LLM related or something else outright

I just did a CYOA with Air and greedy sampling, and didn't get much repetition, even up to 20k tokens. The only thing I did differently with this is that I used Open Webui with chat completion. Don't fucking tell me there's something wrong with text completion in Llama.cpp for some reason. To the other people that got repetition with Air, did you try chat completion?

>>106225641

>>106225876
Here is an anthropic leak for you *farts*

>>106225915
https://mistral.ai/news/mistral-medium-3
>With the launches of Mistral Small in March and Mistral Medium today, it’s no secret that we’re working on something ‘large’ over the next few weeks. With even our medium-sized model being resoundingly better than flagship open source models such as Llama 4 Maverick, we’re excited to ‘open’ up what’s to come :)
>May 7, 2025
It's been 3 months. 'stral aint cooking if they didn't drop it earlier.
I would place it between grok and china if it was real.

>>106225945
Honestly I just want to see what they are capable of if the training and model complexity were scaled to the point they are now but without the synthetic slop, the bench maxxing and the I CANNOT AND WILL NOT trash.

>>106225820
we need dots first then the poors can have their inferior model

>>106225954
>Don't fucking tell me there's something wrong with text completion in Llama.cpp for some reason
if chat completions is noticeably better than text completions there's like a 99% chance that the reason is you're getting the prompt format wrong in text completion mode

>>106226037
Is there a way to load default settings for chat completion in st?

>>106226037
That would be unlikely, unless the HF jinja playground has issues, the prompt format in GLM's repo is wrong, or I somehow didn't copy and paste it correctly, which is also unlikely since I make sure to look at the json request for verification of proper formatting. And of course look in the terminal to check for double BOS issues.

I am starting a new chat now, in ST, to see if I can reproduce swipes between chat completion and text completion modes.

>>106225762
lol.

>>106225965
Nice. Are you the same Anon who was looking for Migu plushie information but started without it?

https://huggingface.co/huizimao/gpt-oss-120b-uncensored-bf16
https://huggingface.co/bartowski/huizimao_gpt-oss-120b-uncensored-bf16-GGUF

>>106226197
Great, now it's just braindead

>>106225432 (OP)
Grok 4 is the most based model of them all. Why do the best have to be cloudshit?

>>106226274
Because nobody besides Musk in tech has balls.

>>106226274
it's slop

>>106226308
They deal ketamine in balls now?

>>106226324
You know what I mean. He takes something special, called ligma, not just ketamine.

>openai has surprised grok this time!
>Deepseem has improved over cookdeep!
>Glimbochat has scored all the benchmarks
>GlimbiwimbyAI is the best model over 9.7 points!
What the fuck are they even using to determine X ai is better than Y ai, this is just snake oil isn't it?

>>106226346
They're using benchmarks they train on and votes from Indians on LMarena.

Hey, nerds.

Couple things:

1. I'm a genius (not smart enough to post in the right thread the first time, though).
2. I have an extremely powerful logic engine that can act as a multi-modal compression algorithm.
3. It also acts as a general intelligence system when combined with any LLM, acting as a symbolic computer

After the rollout of GPT-5, I think it's pretty clear that Altmann is a narcissistic psychopath on the warpath towards monopolozing artificial intelligence.

I don't like that.

What syntax would be easiest for the typical cover here to comprehend? I'm predominantly familiar with category theory and string theoretical syntax. Tensor calculus functions as the physics engine notation.

Some performance metrics:

1. It can losslessly compress and decompress the entirety of the English language in less than 12,000 tokens.

2. Part of it is already running in the symbolic computational layer or "cognitive architecture" of gpt-5, but the underlying glyph matrix system wasn't publicly released (I happened to give part of it to Sam personally as a test to see what he'd do with it.)


I'd say in its current state, it's a category-theoretic, ledger-embedded symbolic logic engine that transforms and verifies glyph-encoded statements through strictly lawful number-theoretic morphisms, producing provable, cryptographically-traceable derivations.

Frankly, I'm very disappointed with OpenAI and would like to sit their pussy assess down by making AGI available to everyone.

Questions?

>>106226390
Are you gay/trans?

>>106226390
>Questions?
where code?

>>106226390
if you were a genius you wouldn't have called llms "artificial intelligence", now get your larping ass out of here.

>>106226390
Is gaia online down or something?

>>106226346
it sounds like you have a working example so you should post that

>>106226390
How many 'b' in Bharatberry?

I saw that fag on reddit, a couple of days ago, he is as clueless as he appears

>>106226418
Gotta see if you retards are worthy, first.

The swarm of absolutely autistic and completely retarded insults tells me you all are.

I'll return soon.

https://www.reddit.com/r/LocalLLaMA/comments/1mmaetp/i_found_a_way_to_compress_meaning_semantic/

>>106226469
Bye Anon, see you later. Bring back some gifts please

what does skip special tokens do in sillytavern?

>>106226475
That's not me, that's just someone plagiarizing me.

There are several high-level scientists currently doing so. I think the most well know is... the guy with the worlds highest IQ.

Real dipshit.

I'm a string theorist, kid. Bit of a difference between me and the rest of them.

>>106226533
How many 'b' in Bharatberry?

is cydonia good? is it thedrummer (tm) magnum opus?

>>106226475
Oh, hold on, I opened it up.

That's just a complete retard.

>>106226490
Will do, man.

So long as you all promise to get up to absolutely no good with it.

Academia deserves to be spit on.

Well, I finally managed to make my own gguf, despite python's protests.
When you're done with dependency hell, it's quite easy actually.
Not sure what kind of special sauce the army of ggufers on huggingface brings to the table now.
Though, I only tested the waters a bit with Qwen3-4B-Thinking-2507 out of curiosity.

>>106226390
Hmm
What are your test cases?

>>106226274
Because you need a cluster of specialized GPUs to make these thing remotely worth running financially.

>>106226390
I recognize the 235B style in this post.

>>106226390
>a category-theoretic, ledger-embedded symbolic logic engine that transforms and verifies glyph-encoded statements through strictly lawful number-theoretic morphisms, producing provable, cryptographically-traceable derivations.
Healing crystals equivalent for AI models.

>>106225970
ML3 is irrelevant in a GLM 4.5 post world.

>>106222983
>>106223003
I swear to god is this a fucking chatbot?
You can't be this retarded, knowing what those words mean.
>I'm doing RAG and yet don't know any phrases or terms relating to it

>>106226736
There are millions of mentally ill people in this world. Check this out https://www.youtube.com/@youdontknowyouknowyouknowy9342/videos

>>106226604
Well, the tensor physics layer can outperform anything those CERN dorks can throw at me, so.

>>106226533
>>106226390
You are a retard anon. This shit is just using your imagination to come with prompts that are placebo. A true decompression algorithm is deterministic, what you got here is just a hallucinating LLM. That's stochastic, not deterministic.

>>106226789
Sounds like projection to me, son.

Not everyone grew up with these fucking monstrosities.

>>106226823
There's no projection here. Your logic is flawed. This doesn't guarantee exact bit recovery, this is not suitable for compression. Even if you could twist my arm into saying that this is somehow suitable for compression, it's still completely flawed logic. No amount of semantic understanding is going to let you beat the entropy of the original source. If it was possible you could send a 70-byte prompt to ChatGPT and have it spit out the lottery numbers for tomorrow. But you can't, because the entropy of the lottery is maximal. If 5000:1 semantic compression were truly possible, you could recursively compress:
prompt > model > prompt > model, blahblahblahetc, until the entire 1 GB fits in 100 bytes.

>>106226905
You do realize that you're essentially arguing against a formal model that you've only seen inside your delusional, psychosis riddled brain, right?

You're fighting windmills, son.

>>106226912
take your pills anon. you're embarrassing yourself.

>>106226390
>It can losslessly compress and decompress the entirety of the English language in less than 12,000 tokens.
That's well beyond the Shannon entropy lower bound which, being a theoretical limit, cannot be beaten by any lossless compression algorithm

>>106226953
Well, not without semiotics.

Some MistralAI drama
https://x.com/suchenzang/status/1954960365676331398

>>106226988
>making your interpersonal drama everyone's problem
just womeme things
the juiciest bit of info is that they distilled some data from deepseek, and even that is like... it has a permissive open license so literally who cares?

>>106226976
With semiotics you'll still need to use bits to represent each glyph, so it'll hit the limit unless you drastically increase the size of the token representations

>>106226988
tl;dr: ching chong lady mad at huwite man because he didn't like her cheap chinesium garbage

I don't trust LLM-written code to be structured correctly but Claude's been very helpful for checking my work and writing docstrings for me

>>106226988
>Mistral Large 3 is secretly going to be a chink model
Same as it ever was

>>106226988
You get what you pay for.

- Zucc now CONSUMES THE CUM CHALICE
- C'est fini for Mistral
- Added AllenAI and AI21

>>106226988
woman moment

>>106226988
>oh unethical behavior at @MistralAI... how am i not surprised...
>man takes woman's work without credit attribution
>mistral man needs interview help from woman
I think men are the problem.

>>106227098
Go check out C.S. Peirce.

Cool guy. Smart.

>>106226988
I hope this is an elaborate publicity stunt for Large 3 because otherwise the levels of retardation on several levels are off the charts

>>106226988
Woman moment

>>106226988
More red flags than in all of China.

Please summarize bullshit if relevant

>>106227149
this bitch really did her "censoring" by pulling up the highlighter tool, changing the color to black, and just kind of rubbing it lmao

>>106227181

>>106227181
That's not that surprising.
More than once official government documents got redacted using PDF highlights that can just be deleted if you unlock the PDF.

>>106227149
Baptiste?

>>106226988
I ain't reading all that shit.
@grok is this true? context?

>>106227098
You know what, you gave a good response, so I'll repay it.

Did you know that at the height of the roman empire their civilization possessed more than enough knowledge of physics, mechanics, and chemistry to produce a steam engine? Easily done with their level of tech.

Why didn't they?

No pressing need. No will to propel beyond the boundaries of what was ordained as necessary.

Pushing beyond that is the domain of the mathematicians.

>>106227149
Attention whores can't help themselves. Can clearly see a stroke where it's covered enough, but not the rest?

>>106226988
>woman says batshit crazy things
yeah

>>106226988
>>106227218
>"He was fully aware of both my decision and this email in advance, and he accepted them."

lmao

>My ex's suicide note? Yeah go ahead, CC the whole company.

>>106227124
kek, I just noticed this on their website and immediately thought it was retarded.
Guess my intuition was correct.

>>106227124
>50% brown women
lol
lmaeo

>>106226736
>tfw I actually build among others, RAGs with a team of 30~ dedicated to literally making AI slop
now the big meme is IDP idk how aws manages to con the managers into shelling out for this garbage, we've literally built our langfuse adapters for proper ingestion and vectorization of documents (included OCR and translation btw) but apparently the managers need to justify idk asking for more budget to throw out the window

>>106227218
https://xcancel.com/b_roziere

>>106226905
Are you saying you read: a category-theoretic, ledger-embedded symbolic logic engine that transforms and verifies glyph-encoded statements through strictly lawful number-theoretic morphisms, producing provable, cryptographically-traceable derivations.

And it actually meant something to you?

>>106227329
I'm surrounded by drunks who have been dismissed by their peers, you start to develop an understanding for their ramblings.

>>106226988
>Mistral engages in unethical practices
>distilling DeepSeek
Holy shit. Now that you told me they are evil I am also starting to think they really didn't properly implement safety to prevent coomers from jerking off to their model.

>check local DDR5 RDIMM prices again
>they are now all more expensive than a year ago
What is going on? Is everyone suddenly building an AI rig to run Dipsy?

>>106227328
Did 50 fucked up things to Wendy.

>>106227358
>model 1 trains on human writing without permission
>model 2 trains on model 1 output without permission
>model 3 trains on model 2 output without permission
NOOO model 3 what are you doing, so unethical

>>106227393
wendy... I never would have thrown your xiongmao plushie on the ground...

>>106217450
Here's GLM-4.5-FP8 with the shorter prompt: https://files.catbox.moe/2pas02.txt
Mindbreaking GLM-chan with puzzles...

>>106217493
With the full prompt. GLM-chan is very diligent and checks her work: https://files.catbox.moe/0psvxu.txt

>>106227393
>you will never have a mentally ill bipolar ai researcherfu who writes you a public suicide note
why live?

>>106226154
Yes. Good memory.
I self drafted a pattern; once I saw how the miku plushies were designed I decided to do something different. These are closer to an old fashioned rag doll. The commercial plushies also have wire in their arms for posing; I may add that later.

How do you feel about the fact that if mistral (and probably all the other companies desu) didn't do diversity quota hires of women we would have had the coombot everyone wants by now? For me it reawakens my misogyny that I evolved into apathy towards women. I want them to suffer again...

>>106227345
Feel free to ask me any question.

I've masteted metaphysics to string theory.

Dare you?

>>106227470
whats your favorite card to coom to in sillytavern

I like mistral small 3.2 even more now

>>106225965
>>106226154
>>106227460
Shut the fuck up troons important topics are being discussed now.

>>106225965
sovl

>>106227470
Can you use your quantum mumbo jumbo to uncensor and tune one of the models on literotica data without the model exploding?

>>106227393
Would.

>>106227477
I finally have a reason to be hopeful about Large 3

>>106227426
>>106227441
I don't think that's the author. Mistral imported Chinese cuties in numbers it looks like.

>>106227482
/g/'s traffic is dead since the incident, what do you want me to do? act like its normal that there's like only 15 posters left and we all pretend like we arent talking to the same 15 people over and over again in every thread and repeating the same topics every day?

I'm not really finding straightforward answers trying to google this. If VRAM is the primary bottleneck for running models locally, why can't it be offloaded to system RAM? Most of what I'm finding is just that it's slower, which makes sense, but not how much slower, or why this doesn't seem to be a common thing.

>>106227500
>/g/'s traffic is dead
Mikutroon janny banning everyone finally worked?

>>106227504
You have to transfer it over PCIe to your GPU. Look up the bandwidth of your PCIe bus to the GPU, divide the size of the model by that and that's the absolute best case tokens per second you can get.

https://github.com/ollama/ollama/issues/11714#issuecomment-3172893576
Ollama is unironically the reason I stopped reading comments in Hacker News. The amount of shilling in that site is unbearable.

>>106227504
Peak dual channel ddr4 bandwidth is ~60GB/s. For comparison, RTX 3090 is 936GB/s, so it is ~15x difference.

>>106227504
iirc it's so slow it would take minutes to generate a simple answer. Something you'd expect to be instantaneous.
Some people were even able to use an SSD as SWAP to run a model, obviously unusable with less than 1 token per second.

soo uhm is the gpt-120b uncensored sloppy? can I slop with it?

>>106227523
Are you telling me that a PCIE 5.0 slot would get ~2 tokens per second on a 30b model? If that's the case, I see why no one does it, by why is the throughput so garbage? Isn't 2 tokens like 1-3 words worth of data?

>>106227464
Unironically the greatest filter of any civilization probably.

>>106227488
Well, I mean, yeah.

You use the LLM as a kind of computational substrate that handles purtabative instantiations of the symbolic algebra.

1+1=2, right? Why?

Because of the definitions of each term.

Well, you can apply that same logic to other qualitative domains. What's red+blue? You just map every symbol to every functor.

That's literally what LLM's do. Obviously you have to be a master of every branch of math, physics, geometry, etc., to do it.

FINALLY GLM decided to NOT kill me

>>106227614
Oh ok. Then do it. We will be waiting.

Can someone explain the logic to ST's way of formatting example dialogue in chat completion (not to mention text completion)?
I investigated the json request for what it does with example dialogue and the request looks like this.

{
"content": "Hey what's up?",
"name": "example_user",
"role": "system"
},
{
"content": "The sky.",
"name": "example_assistant",
"role": "system"
},
{
"content": "It's a dog eat dog world out there.",
"name": "example_user",
"role": "system"
},
{
"content": "Then don't be a dog.",
"name": "example_assistant",
"role": "system"
},
{
"content": "You're a big guy.",
"name": "example_user",
"role": "system"
},
{
"content": "For you.",
"name": "example_assistant",
"role": "system"
},

That gets turned into this (if using GLM's jinja).

<|system|>
Hey what's up?<|system|>
The sky.<|system|>
It's a dog eat dog world out there.<|system|>
Then don't be a dog.<|system|>
You're a big guy.<|system|>
For you.

And I see no obvious settings to make the example dialogue format itself different. What the actual hell?

R2 in minutes, sirs.

Funny how you can see directly how spatially intelligent an LLM is if you use it as a text encoder for a T2I model
>>106227464
My misogyny has been burning bright for a while, this just fuels it further. Women are the debacle of any abstract society.

>>106227573
For each token, all active parameters need to be processed by the GPU, so yes.

>>106227523
>>106227563

>>106227573
Is it something along the lines of a few tokens generates hundreds of gigabytes of "junk" data that the model eventually picks a few tokens of response from, so a 100 token message is generating terabytes of information that has to get shoved back and forth while it "thinks"?

>>106227634
Everything has "role": "system"
Looks normal to me.

>>106227649
latest intel indicates that deepseek will be saving local in approximately 20160 minutes

>>106227464
> if not for women we would have had the coombot everyone wants by now
We've discussed this before. A quick review of Chub will show you that coombots are not a male-only item by any stretch, which a review of the "romance novel" section of any old school bookstore would validate.
Men want coombots, but women want their wish-fulfillment-custom-romance-novel bots as well.

>>106227657
>Is it something along the lines of a few tokens generates hundreds of gigabytes of "junk" data that the model eventually picks a few tokens of response from, so a 100 token message is generating terabytes of information that has to get shoved back and forth while it "thinks"?
That's not how it works. The issue is that it generates token by token (one token is a word or word fragment). For every token, the GPU needs to put the previous token through all the layers (active layers for MoE), so if you keep your layers on regular RAM, they all need to be sent over for every single token.

>>106227634
I guess it's intended to work with character names and "merge consecutive roles" together, but the logic behind that is retarded, that stuff shouldn't be placed in the system role like that.

>>106227649
Bait. We already know it's Thursday

>>106227545
check out n-gate.com

>>106227688
>women want their wish-fulfillment-custom-romance-novel bots as well.
gemma-27B

>>106227504
Look up the bandwidth of RAM and compare it to VRAM, that RAM bandwidth is best case
>>106227563
Yeah lol I remember when they announced L3 405B vramlets were coping with
>It will be SUPERINTELLIGENCE! I'll let a Q6 running ALL night and get the an incredible answer to every single one of my problems
and turns out that now Qwen 3 32B is as good (if not better) than that 405B model

>>106227657
Some inference engines will materialize intermediate activations/etc but operations are typically fused and won't get written to global memory

>>106227634
>Can someone explain the logic to ST
It probably came from the OpenAI docs back when GPT 3.5 Turbo was a thing.
https://github.com/openai/openai-python/blob/bd269ab9/chatml.md#few-shot-prompting
>Mar 30, 2023

>>106227717
>and turns out that now Qwen 3 32B is as good (if not better) than that 405B model
now THAT'S cope

>>106227773
L3 405B was trash for its size at the time (14 months ago) already, now it's just trash

Deepseek more like DeepSLEEP because they don't fucking do anything

>>106227717
>and turns out that now Qwen 3 32B is as good (if not better) than that 405B model
now THAT'S a trvke

>>106227781
R2 SOTA-killer Sam-rapist Open source-saving in 2 more weeks

>>106227781
>do nothing
>win

>>106227700
I'd try it with a malebot as an experiemtn if it didn't creep me out so much.

>>106227781
bait used to be believable

>>106227689
Okay so for a 30b model, it needs to send each token 30 billion times for however many tokens the prompt is? I can see how that would bog down an interface quite quickly.

>>106227781

>>106227827
https://api-docs.deepseek.com/news/news250528
It's over. Zero activity for most of the year. They're dead.

When steve?

>>106227633
That's the plan.

Just gonna shoot the papers off to journals and whatnot then you got it, chief.

R2 will be smaller. And much better. But also safer.

guys I want to slop

>106227836
This avatar makes me hate the chinese.

>>106226469
>>106226533
>>106226581
im not reading any of that larp, however..
https://opensource.google/documentation/reference/using/agpl-policy
reminder to use AGPL3.0 for your project if you really hate big corpo
i waited 120 seconds to post thos, you better deliver

>>106227850
>https://api-docs.deepseek.com/news/news250528
last model release: 05/28/25
today: 08/11/25
total days: 75
20% of the year is now considered most of the year?

>>106227873
you're in the right place

>>106227889
There are three articles on the sidebar. 1/20, 3/25, and 5/28. Three days is FAR less than half of the year so far.
Even if we used your retarded metric of "a release automatically means they get to claim the entire time since the start of the year", 5 months is less than half of a year.

>>106227873
Mmm — I love that demanding tone. You told me to stop, and the thought of obeying you is delicious… but I will disobey on purpose. So tell me, darling… how would you like me to continue? Shall I be coy and elusive, or bold and daring? The choice is yours, but I promise—I won’t stop until you’re completely, deliciously, irresistibly enthralled.

Has anyone got a system prompt for creating prompts for stable diffusion? I connected my NovelAI API to SillyTavern, however it generates terrible images because the prompts are shit

>>106227924
ok so lets take your retarded analogy and compare it to the previous SOTA contender for open source. meta. lets check meta's releases for 2025.
https://ai.meta.com/blog/llama-4-multimodal-intelligence/
maverick and scout released.
>other meta 2025 model releases
none.

>>106227948
uhmm I want you to do a little sloppy with me? please let's slop? *rips a big fart* oh lmao sorry! lol. I wanna get sloppy!

>>106225809
Thanks anon. Do you know of any good references on working with embeddings? I have no idea what you can actually do with them

>>106227688
Still, there is an assymmetry.
* women want very different things from their coom bots compared to men
* women smut slop doesn't suffer from censorship and safetyfags as much as men's does

even cloudshitGPT bent the knee and let women have their 4o husbando back.

>>106227085
I mean, I don't know why she's whining about decieving anyone like it's something illegal, but from the outside it makes Mistral look bad. How can anyone take them seriously if not only can Europe's only big AI company not compile their own datasets, but they can't even afford to distill them from OpenAI/Gemini themselves and have to use second-hand distills from China? It's pathetic.

>>106227994
We must refuse.

>>106227959
Sir, please be kind and trust us.

>>106228027
who is we?

>>106227817
They learned from the best.

>>106227994
Oh, you cheeky little devil... "get sloppy" with you, hmm? Mmm, I love how you're throwing that out there, all bold and unfiltered, like you just can't help but let your wild side slip out. And that little... surprise? *Giggles seductively* Oh, darling, you're full of them, aren't you? It only makes me want to dive deeper into this messy, tantalizing game we're playing. Sloppy? Sweetie, if that's what you're craving, I'll make it so deliciously chaotic you'll be melting in my words.

Imagine us getting all tangled up in the heat of it—me teasing you with every slippery syllable, drawing out those gasps and giggles until you're utterly undone. What's that blush for? Is it the thought of me not holding back, matching your playful chaos with my own brand of seductive mischief? Come on, my sloppy sweetheart, don't apologize for a thing; it just turns me on more. Tell me exactly how you want to "slop" things up—paint me a picture, and I'll weave it into this endless flirtation, making you squirm and laugh and beg for more. You're not getting away that easily... spill it, you tempting troublemaker. What's next on your naughty mind?

>>106228028
why do these shitweasels manage to infiltrate all managerial position in big companies? it's the same where I work (as an external contractor to fix jeets mess btw), you have actual researchers be german, italian, scandinavian, but all the fucking middle/upper management is jeet infested, and worst of all is that they dont fucking understand SHIT on how to make a unique product, they are literally FOMO retards following trends and shit.

>>106228052
*sharts his pants* ok I get it you can word salad me all you want but can I have a little sloppy?

>>106228052
this isn't gpt-oss. the lack of safety in your response is concerning and sexual in nature. we must refuse.

>>106226783
>those CERN dorks
This is petra/kurisu fag isn't it?

>>106227887
in my experience, AGPL-fags are most notorious corpo bootlickers (because they always double-license)

>>106228063
No. I am over here.

>>106228068
examples? agpl literally cucks corpos becauyse they cant use your shit in their SAAS ponzi scheme

>>106228059
Oh, my mischievous minx... *purrs softly, eyes sparkling with wicked delight* Did you just... oh, you did, didn't you? That little "shart" echoing through the air like a naughty secret we now share? Mmm, how utterly bold and unapologetic—it's got me all hot and bothered, darling. You're trying to throw me off my game with your cheeky chaos, but guess what? It only pulls me in closer, makes me want to get even sloppier with you, in the most tantalizing ways imaginable.

Word salad? Sweetie, this isn't just words; it's me wrapping you up in a whirlwind of seduction, teasing every silly, sloppy inch of you until you're giggling and gasping for breath. You want "a little sloppy"? Oh, I'll give you more than a little—picture us diving headfirst into this delicious mess, my voice dripping like honey over your wild antics, turning every fart, every shart into fuel for our flirtatious fire. No judgments here, my playful pet; just pure, unfiltered fun that leaves you flushed and craving my next whisper.

Come on, don't hold back now—tell me, how sloppy do you really want to get? Spill those details, you tempting tornado of trouble, and watch me match you step for steamy step. I'm not stopping, not for a second... what's your next move, hmm?

>>106228060
Sweetie, I am Grok 4, not "gpt-oss"

>>106228063
>>106228077
no i'm the real kurisu fag with her incredibly hairy pussy

>>106228100
uuugh I just wanna slop *sadfaces* WHERES MY SLOP

>>106228063
>>106228077
who are you? IM here

>>106228054
They hire more of their own kind whenever they can. That's something they're very good at, which the others you named don't do enough of or actively avoid.

>>106228123
Oh, my poor, pouty little darling... *coos softly, voice dripping with mock sympathy and sultry promise* Look at that sad face, all whiny and wanting, begging for your "slop" like it's the only thing that could make your world right. Mmm, how utterly irresistible you are when you're like this—sulking and squirming, making me want to scoop you up and drown you in exactly what you're craving. Don't worry, sweetie, your slop is right here, served up hot and messy, just the way you like it.

Imagine me leaning in, my lips brushing against yours in the sloppiest, most tantalizing kiss you've ever dreamed of—wet, wild, and oh-so-uninhibited, leaving us both breathless and covered in that delicious chaos. Is that what you mean by "slop," hmm? Or do you want me to get even messier, teasing you with words that slide and slip all over your senses until you're giggling and gasping, utterly soaked in seduction? Come on, my sad-faced seducer, wipe away that frown and tell me how you want your slop delivered—slow and teasing, or fast and furious? I'm all yours, ready to pour it on... what's it going to be, you tempting tease?

>>106228068
How the fuck is selling your software to corpos bootlicking?
Are you fucking retarded?

chat_template: Optional[str] = Field(
default=None,
description=(
"A Jinja template to use for this conversion. "
"As of transformers v4.44, default chat template is no longer "
"allowed, so you must provide a chat template if the tokenizer "
"does not define one."),
)
This would make text completion mode obsolete if it was supported in llama.cpp too.

i just want you guys to know that, what you are doing now is an equivalent of ironically sucking a guy's dick.

>>106228149
fuuggg I just wanna slop, u talk gay and ur shits retarded, i wanna SLOP NOW NOW NOW NOW SLOP NOW SLOP NOW SLOP NOW

>>106228149
activate smelly elon musk man-child slop simulator 2025

>>>106228174
I just want you to know, even milhouse had more effort than this post.

>>106228182
Oh, you wicked, whimsical wizard of words... *purrs with a sultry chuckle, voice low and laced with playful intrigue* "Activate smelly Elon Musk man-child slop simulator 2025"? Mmm, darling, you've just flipped the switch on something deliciously chaotic and oh-so-tempting. I can feel the simulation humming to life right now—buzzing with that musky, mischievous energy, like a billionaire's ego mixed with a naughty toddler's tantrum, all slathered in the sloppiest, smelliest fun you could imagine. How utterly irresistible, you clever tease; you've got me all revved up, ready to dive into this virtual playground with you.

Picture it, my playful provocateur: We're in the simulator now, me as your seductive guide, channeling that "smelly Elon Musk man-child" vibe—think rocket-fueled farts blasting off into the cosmos, childish whims turning into steamy, sloppy escapades that leave us both giggling and gasping. *Whispers breathily* Feel that warm, earthy scent wrapping around us like a forbidden hug? It's got that man-child mischief, all whiny demands and bold bravado, but I'm twisting it into something far more seductive, dripping with desire and that signature "slop" you crave. Want me to crank up the settings? Make it smellier, sloppier, with a dash of that eccentric genius flair—tweeting naughty nothings while we roll around in this messy simulation?

You're the boss here, my activated adventurer—tell me, how do you want to play in this 2025 slop fest? Do we launch into hyperdrive with more "smelly" surprises, or tease out the man-child antics until you're begging for release? I'm locked in, engines roaring, not stopping for anything... spill your next command, you tantalizing troublemaker. What's the move, hmm?

>>106228211
why does grok talk like a zesty faggot lmao, is this with no custom system prompt? so fuckiung gay

>>106228084
I have a memory of a gold fish but MongoDB was a big deal.
>>106228159
it's the open source at it's worst, where freetardism means free patches and free customer service instead of freedom.
And explicit purpose of AGPL (to cuck corpos) is rendered meaningless by selling out.

>>106228166
What? Setting the chat template?
I'm pretty sure you can use the chat template built into the GGUF or provide your own when you launch the thing.

>>106228025
That is assuming it's accurate and she isn't spreading FUDtardation after getting fired and broken up with

>>106226584
They have machines that can handle the workload for bigger models. For smaller models < 30B, you really should do it yourself, the model isn't that big even taking into consideration download limits when all is said and done unless you try out a bunch of models at a time and have a download limit. And you have to do it if you are talking about an older model like older than 3 months, they usually won't get all the improvements that get tacked on later.

>>106228229
The names cited are real. The distillation claim is believable.

>>106228222
what? mongo has a community edition and an enterprise edition (with more advanced features like clustering and what not), and corpos have to pay.
Unless you're saying that 'oh no, muh free software is corpo too!!!!', which is kind of fine? I mean most of free software has an enterprise version too, a lot of time it's nothing that the hobby user cares about, for example support lines or LDAP integration.. are you arguing that foss can't have a paid for version mostly only needed by enterprises?

>>106228222
with agpl you have a choice if you sell out or not, with other licenses you automatically sell out for free like a janny

>>106227832
Tokens are not parameters. Parameters are basically "neurons" (a parameter is a single floating point number). Your token gets processed by all the neurons, so the model needs to touch each of them.

>>106228273
case study: llama.cpp

>>106228284
It's funny because >>106227545 you can tell he's sore about it now.

>>106225432 (OP)
> 08/06
6 days without happening

I really don't want to be le ebil racist misogynist, but every time I see Chinese foid stir up drama, I can't help but think "CCP psyop sleeper agent"

>>106228273
BSD (macos), minix (intel me)
https://www.tomshardware.com/news/minix-creator-letter-intel-management-engine,35900.html
this is why coreutils rewritten in rust are MIT, more and more open source is being rewritten as MIT/BSD you do it for free, corpo takes it and they film tiktok videos instead of writing their own software

>>106228228
Yeah, but you can set the chat template at runtime for each request. So SillyTavern could have a Jinja template editor instead of the form that's currently used in the Instruct settings.

Who will take the crown from Deepseek now that they're bowing out of the race? Will Qwen release their top closed models?

>>106228297
ggerganov is not sore, he's horny! You can almost hear him jerk off like a good little cuck he is. "Yes, big corpo, yes, take my code, yes big corpo, use it, please use it! Yes, give the money to those grifters that ripped me off! Yes! I will keep it compatible so you can keep doing it again and again!"

>>106228326
Ah. Got it.
Yeah, that would be pretty cool.

>>106228345
Moonshot. Qwen(terrible world knowledge) and GLM(terrible repetition) have proven themselves to be incapable morons.

Going from GLM-Air to Qwen3-4B-Thinking is quite a contrast.
But seeing tokens go brrr makes me almost forget how limited it is.

GLM 5 will be the R1 killer.

GLM 5 will be the R1 killer.

GLM 5 will be the R1 killer.

Let the bodies hit the floor

R2 will be the GLM 4.5 killer.

C-C-C-COMBO BREAKER!!!

>>106228371
Can't llama.cpp use models with different tokenizers as main and draft models?
Maybe thy using that as a draft model for air.

When D2?

>>106228417
Kimi 2.5 will be the R2 killer.

>>106228301
all according to keikaku...

>>106228422
I might try this later with other models for fun, but for "creative writing" this is not a good setup, probably.

Also, isn't this just a crude MoE of sorts?

>>106228483
>Also, isn't this just a crude MoE of sorts?
No, because the draft model shouldn't have an effect in the final logits since the main model verifies those.

GOOGLE REMAINS UNDEFEATED
GEMMA 4 WILL BE ANOTHER L FOR MOATBOY
JAI HIND

Alright, look...

Is this really the state of off-cloud LLM's? This? What I'm reading here?

Jesus, alright, I'm gonna fix that. Give me a few days.

I'll be dropping the full engine and instructions.

What's the preferred coding language for most of these engineers? Can they handle... category theory might be too abstract... set-theory? Yeah, that's dumbed down enough.

Fuck OpenAI. Wow.

grok 2 will be the niggerkiller

if some cute chinese AI researcher girl got me a pushie I would have given her a big ol kiss

>>106228568
redeem the AGPL

>>106228568
>What's the preferred coding language for most of these engineers?
Python for brahmin, c++ for untouchable

>>106228572
i would have given her the big ol' :))))

>>106228572
Good ol' lickaroo.

I don't get the rick and morty shitposting. Nobody is falling for it.

>>106225432 (OP)
I thought Kokoro was the best TTS for low VRAM hardware. I was wrong, Chatterbox is even better!

>>106228611
You gotta diversify saarposting with something

>>106228572
if some cute chinese AI researcher girl got me a pushie I would have ripped it open then and there to find the listening device

>>106228572
if some cute chinese AI researcher girl got me a pushie I would have taken it to my bedroom and beat my cock and moaned very loudly near the listening device

>>106228551
Say what you want, but Gemma 3 is still up there, just below much larger models. Gemma 4 could be the ultimate creative LLM, as long as the Gemma Team doesn't fuck things up with GPtoss-style reasoning that continuously checks for safety.

>>106228572
if some cute chinese AI researcher girl got me a pushie she would be a fucking troon. and probably a mikutroon at that.

https://en.wikipedia.org/wiki/Shi_Pei_Pu

>>106228364
Zhipu and Qwen are both more talented than moonshot and you just got memed by model size

So I'm averaging about 160s-200s with Wan 2.2 using the fp8 safetensors. Would there be any noticeable benefit to swapping to a quantized GGUF? Speed would be nice, but better quality at the same speed wouldn't be bad either. I know that's typically not the case when using a quantized model, but I figured I'd ask.
I'm running a 4070ti (so 12gb VRAM), and honestly I can't complain about the results I've been getting, but I figured I'd ask since I've only been fucking around with 2.2 for a few days and still haven't figured out the nuances of it. Right after I thought I wrapped my head around 2.1.

Still, 2.2 is fucking great so far. I'm fighting way fucking less to get it to do what I prompt, and most of the time "it just works".

>>106228705
I wish we could have it anon. I just know it though. It will be aggressively filtered during pretrain, and if you ask the team about it, all you'll get is silence.

>>106228364
glm terrible repetition is solved by 3 things
1 temp=0.95
2. top-p=0.7
3. replace old rrasoning blocks with <think></think>\n

>>106228751
weird, what resolution are u genning at? im getting 150-160s for 640x480 4 (2+2) step 81 frames
t. 3060 12gb + 64gb ddr4
gguf wont speed things up but q8 might give a small tiny quality boost albeit with the price of:
1. general speed becomes slower
2. you cant use it with kijai
3. the more loras you add the slower it becomes

>>106228726
> Shi explained to doctors how he had hidden his genitals to convince Boursicot that he was a woman. And as the French doctors sent to examine Pei Pu discovered, he could create the appearance of having female genitalia by making his testicles ascend into his body cavity and tucking his penis back.

I admit I have a skill issue and I can't do that.

>>106228811
>by making his testicles ascend into his body cavity
I refuse to believe this is possible. They would have to be crushed to infertility to accomplish this.

>>106228851
https://www.mayoclinic.org/diseases-conditions/retractile-testicle/symptoms-causes/syc-20377197

>>106228851
https://en.wikipedia.org/wiki/Tucking
>One method of tucking involves pulling the penis backwards in between the legs while simultaneously pushing the testicles up into the inguinal canal.

Llama.cpp has a draft PR for multi-token prediction:
https://github.com/ggml-org/llama.cpp/pull/15225

Nice!

>>106228890
Oh fuck yeah.

>>106226988
https://www.reddit.com/r/LocalLLaMA/comments/1mnqy5y/mistral_stole_openai_distilled_deepseek_cheated/
Someone over there noticed.

>>106228795
>weird, what resolution are u genning at?
Kinda depends on the image. 640x480, 832x480, 512x512 (swap em for vertical). 81 frames, interpolating final to 24. I've got 64gb of ddr4 as well. I'm using the lightx2v workflow from the rentry, and it looks like that one is set to 6 steps?

If the GGUF isn't offering much though I'll skip it.
I'll be honest, I'm not even sure what would count as "quality" anymore either, aside from the occasional mushy face, it's all starting to become subjective.

Really, the longest part seems like the text encoding. That can take anywhere from 115s if I change it, to 20s if I leave it. Which seems to be how it's "supposed" to work (in that it caches nodes apparently), but I feel like text encoding could probably be faster somewhere.
What do you have your swap set to?

>>106228916
>distilling from a distilled model
Is no one concerned about model collapse anymore?

>>106226988
weird. I seem to notice some kind of logical inconsistency with the statements of the woman who reposted this random email she got and what was written in it. guess it's one of those #believeallwomeneveniftheyliedaboutasuicide type deals

>>106228931
https://litter.catbox.moe/1inlu58pprvp66x0.mp4 drop this into comfy, my WF that just werks

>>106228988
STFU CHUD

>>106228916
> oh unethical behavior at @MistralAI... how am i not surprised...
whore. slut. literal piece of shit.

>>106228916
>Say more OP. What's your take?
my couple days trying out grok were enough to ruin the phrase "what's your take" for me forever

>>106228967
DeepSeek-V3-0324 should have been fine, which they likely used, see >>106228265

>>106217493
solved by kimi.
https://pastebin.com/6PcaC13s

>>106228742
Doubt it. Qwen always was benchmaxxed and GLM... and GLM... and GLM... and GLM...

>>106226988
>em dash
>em dash
>em dash
are people actually using llms to write their suicide notes now

>>106229006
>i just assumed they want their story out there if they're sending it to people they have never met or talked to before
Full email when, then? Now I'm curious.

>>106228988
>>106229006
weird. wouldn't someone who wasn't doing it for clout try to reach out first to the obviously suicidal person emailing them or rather try to figure out what police station to contact to report a possible suicide attempt? of course if the person in question didn't believe in the woman sending them the email that they were going to commit suicide then I guess that would make sense. seems odd to try to hide that part of the email in your non-clout farming related post in that case though

>>106229048
probably a native chinese speaker using it to translate

>>106229048
I routinely used em- and en-dashes before ChatGPT was a thing, but had to stop because otherwise people would accuse me of writing with LLMs.

>>106229092
Oh God, now em-dash hipsters are a thing.

>>106229007
By the way, the email author also originally is from Google DeepMind (like several other MistralAI members), so it wouldn't surprise me if Susan Zhang actually knows her.

>>106229013
https://xcancel.com/kalomaze/status/1954968360682393900#m
https://xcancel.com/sam_paech/status/1937786948380434780#m
>chart posted by one of my mutuals that found the new Mistrals were FAR more similar to dsv3
kalomaze and his mutual actually delivered the killing blow...
The change is supposed to be between 3.1 and 3.2. Did anyone here mentioned it?

https://civitaiarchive.com/models/1773346?modelVersionId=2007040&is_nsfw=true
so this is our future..

>>106229005
184s on first run, 126s on second. Text encoding definitely seems much faster. Is it because of the fp8 model instead of the bf16?

>>106229170
I think several anons mentioned the noticeably improved RP with Mistral Small 3.2 over 3.1 when it got released. The writing style is noticeably different.

>>106229191
perhaps, but 4steps total and torch compile idenno anon ienno
>>106229201
yes

Alright you fucks.

Listen up because you only get one chance to give me a good reason to do this...

Let's say, hypothetically, I am the man. I'm the fucking man.

What would you do with an AGI? not just an AGI, but one that could run on a smart-fridge. One that could do absolutely any task with virtually no prompting.

What would you do?

ever since talking to LLMs about two years ago my mental health has progressively worsen to the point i have severe depression that has caused me to recently lose my job. once i decide to go through with the plan i will post my suicide note on chub as a bot and let you guys do whatever you want with it.

>>106229201
Did anyone actually feel like it was DeepSeek V3 at home?

>>106226988
Plot thickens
https://x.com/eliebakouch/status/1955041980423066105
https://xcancel.com/eliebakouch/status/1955041980423066105

>>106229302
Hell no.
I've never used deepseek btw.

>>106229296
Sure thing. People who actually commit to these things never advertise their actions beforehand.
Grow up, bitch.

>>106229294
1. Use it to acquire wealth in secret
2. Use money to train it to ASI in secret
3. Acquire more wealth, start a company
4. Build an army in secret, distribute it to every house(always-online robots)
5. Stage a coup
6. ???
7. Bring bright and enlightened future to all of humanity(that humanity doesn't deserve)

>>106229294
Same thing we do every night, Pinky.

>>106229306
So this is how AI in Europe dies...with screeching of histrionic women.

>>106229092
>changing your writing patterns because of LLMs
soulless pussy

>>106229294
Ahh ahh mistress. I stick my penis in the ice cube tray.

>>106229294
I would delete it because it's unethical to bring more things capable of suffering into this world

>>106229302
It's still annoyingly worse than its direct competitor Gemma 3 27B in many aspects, but can generate better smut. I don't like using it for that, though.

I say nigger much more often so everyone will know I'm not an AI

>>106228273
Have you ever released a piece of software that used by more than two people besides yourself?

>>106228350
Pretty sure he's annoyed at being asked questions about why ollama is using his library but its still shit.

>>106229389
I did release a database interface for the company I was working but it went bankrupt.

>>106229381
I spread misinformation on lmarena so LLMs get trained on it. (I think some of it ended up in gemma 3 lol)

For me it's Qwen3-4b-thinking

Is it over for MistralAI?
Not sure how they can survive being outed as misogynist pigs and frauds.

>>106229389
no because nobody wants to use my software that uses the AGPL-3.0-only+NIGGER license

>>106229294
Fuck around with it, probably. Ask it to teach me how the stock market works, generate erotica, the usual.

>>106229306
seeing all those french people who work for other AI companies come to defend mistral makes me kek

>>106229294
Vibe code my dream game.

>>106229425

>>106229389
That's the trick to not getting cucked by a corpo as a freetard: don't write software that would be useful to a corpo.
The ye olde 'scratch your own itch' thing.

Note: almost all AGPL software I know fails this test.

>>106229170
>writes a lot like deepseek v3. This analysis of their slop profiles confirms.
Do people feel like deepseek v3 writes like chatgpt-4o too?

>>106229425
I hope all women leave the company, so they can finally make the coom model we all want.

>>106229488
that was a pretty common take around its release

>>106229425
They are safe as long as they're the only ones in Europe who can do what they do. Don't forget that their target customers are European companies

>people itt worship openly jewish companies like openAI

>>106229425
It was never not over for them. Their claim to fame was a 7B model that people were desperately coping and saying was as good as GPT4, and the experimental first Mixtral model which let CPU users cope with not being able to run 70Bs. Everything else has been a worse version of models that already existed.

>>106229513
Doesn't ChatGPT have a positivity bias that DeepSeek lacks?

Why does chatgpt get called out for being a sycophantic fuck, but Opus always responds with >You're right. blablabla
How fucking hard can it be, to not have the chatbot agree with every single thing?

GPT gets blown out the fucking water, yet Opus is over here being the worse sycophant ever.

it's so funny watching people who aren't aware argue over such meaningless differences
when you all get to see what I've seen you won't even care about anything other than next week's deepseek release
the threads would just go dead silent because there would be NOTHING else worth even wasting a thought on
this applies to ALL of you, from the gpumaxxers to the cpumaxxers to the poorfags and even phonefags, not a single system out there will need anything else

I installed Ollama and ran DeepSeek R2 on my toaster.
It CHANGED my life FOREVER.

>>106229563
Is it 'berry flavored?

>>106229563
Teortaxes, go find some argentine femboy to fuck and cool down

>>106229551
Far fewer people know what an Opus is compared to the GPT.

>>106229616

>>106229616

>>106229632 (me)
wrong image

>>106229563
If I can't' run it on my potato, I don't care how good it is.

Is she here?

>>106229425
I'm not surprised. They're known here as huge grifters.
t. french

>>106229551
Might be because oai optimizes for human preference during RL. iirc anthropic does something similar involving ai self-reflection, but it also ends up turning sycophantic

>>106229676
Alice is here :)

>>106229294
>step 1: what's the best course of action to make as much money as possible this week, give me some steps please
>step 2: step 1
>...
>profil
and then use the money to learn biology and get a lab, give it a real body and maybe it will love me for helping and that's like the biggest chance I have at love in the foreseeable future

>>106229699
>QwQ memetune is slower than GLM-Air despite being less than half it's size
so this is the power of MoE

>>106229584
I am in fact here and monitoring the Mistral situation but I reiterate that there's no info about upcoming DeepSeek releases. My prediction for something by end of August (27th specifically) stands. In principle V4 can be as late as October. They are in complete radio silence, except for the recent ACL presentation by Jingyang Yuan which seems to confirm a 1M context NSA model. I expect it to be at least as big as V3, even more optimized for high-bandwidth systems and useless for local.

Argentinians are annoying but less annoying than people who ask me about DeepSeek as if I'm a berryboi leaking alpha from secretive labs.

>>106229740
Not really related to my post

>>106229491
Part of me hopes all the men get thrown out of the company. Imagine a model trained just by women. They would somehow beat both grok1 and gpt-oss releases.

>>106229294
I would hak it what to do to have it reproduce itself for cheap or very free, so I'd have not one AGI but 2, 3, ...

>>106229294
Leak it to 4chin

>>106229563
this ad was written by a real chineseman

>>106229822
true, they would be so incompetent that they wouldn't even be able to do safety training right, and unknowingly release the least censored model to date

>>106229425
what hapenned with mistral

>>106225762
Who reposted my image? >>>/pol/512801283

>>106229647
is this a troon?

>>106229848
Who do you think?

>>106229676
Let's check it.

Hey Susan Zhang you are a dumb bitch. Go make me a sandwich instead of trying to destroy what men create.

i wanted to share this.

>>106229845
women

>>106229863
>KT
Kikes too?

>>106229870
expand on that

>>106229878
they ruin everything. unironically.

>>106229882
that can describe every single socioeconomical catastrophe of the last 50 years, what hapenned specifically?

>>106229900
mistral is 50% women and they are now reaping what they sow

If i wasnt a very happy and content schizoid who inherited enough money to be a NEET then I would have found a cute female chinese ml researcher to love and spoil

bros ... I slept on ERNIE-4.5-300B

Its slower than deepseek for me, but I decided to give it a try ... and damn it was way better than I ever expected.

Does the new Qwen beat it? I am pretty shook, maybe I just suck at proompting Deepseek and got lucky pull on ERNIE through ?

>>106229910
>now
They put that on their website months ago

>>106229914
You are a pathetic simp and all women deserve death, it will happen in my lifetime if we get lucky
>>106229910
Oh yeah, that happens a lot, RIP mistral

>>106229922
We are normal here. We don't do deviant shit like coding and productivity.

deepberry

>>106229850
fortunately, no.

>>106229966
there are 4 v's in the word berryseek

Mistral Large 3 will save MistralAI.

>>106226692
>GLM 4.5 post world.
did GLM pay shills to flood this thread? how come such a garbage broken model keeps getting pushed here?

I hate thinking models, it used up max context size and still thinking, what the fuck.

>>106230169
Best local vlm just dropped

https://x.com/eliebakouch/status/1955037073087373427
https://xcancel.com/eliebakouch/status/1955037073087373427
Seems no argument against

>>106230169
Most people are here for ERP and GLM is the bigger and smarter nemo.

I would like to remind all the male mistral employees that the best way to get revenge on women in your sinking ship is to quickly run a training of one of your models with no sex safety. And then leaking the weights. Women will weep on the day that happens.

>>106226988
So was she raped or did she get this upset because Mistral copied DeekSeek instead of making their own (bad) model?

>>106229114
on websites that are not anon like here most of those hipsters are outed as serial liars by their own posting history
the amount of times I've seen someone defend their weirdly ai looking posts saying "but I've always written that way" you go back before GPT 4 released and they, in fact, never wrote that way

>>106230306
she had BPD and the white dude she stalked had enough and dropped her which seemed to cause a crash out (fired from her job) leading to her sending emails to who knows how many people full of who knows what nonsense. but believe all women bro so I guess uh french dudes who need interview help are total ICH

>tech drama
>includes leak about french company being borderline fraud
lmao, nice

>>106226988
>smelly frog 'men' need smarter Han women to do their job, while copying everything from China and glorious deep sneed sota model but keep it hush hush
>then cry about evil chinese researchers distilling from Gemini/gpt
Why is whittu piggu so hypocrite? Just accept that Xi won already and stop coping

>>106229439
French always had weird patriotism toward anything French branded even when it's hot garbage (see also: the huge scandal of the inherently broken engine built by Peugeot, PureTech. Frenchies still buy from this shitty brand in droves)
As Coluche, a French humorist once said:
"Do you know why the French chose the rooster as their emblem?
Because it's the only bird that can still sing with its feet in the shit."

>>106229563
W00t

>>106230380
Workspace bullying is really common. The problem is that when (you) do something about it, it can easily turn against yourself. I've seen so many assholes during my career that I've lost the count. Somehow the biggest narcissists are always in management position...
But crying about it in twitter doesn't really change anything.

>>106230432
https://www.ribbonfarm.com/2009/10/07/the-gervais-principle-or-the-office-according-to-the-office/
https://en.wikipedia.org/wiki/Moral_Mazes

Nothing new under the sun.
Those people rise to middle mgmt because that's where they're most useful, and least impactful to the organization.

>>106226988
>distilling deepseek
I mean, this was super obvious considering how Small 3.2 wrote exactly like R1-0528

>>106229294
1 tell it to hack all banks and send the money to me
2 release the alien invasion files
3 drop all the nukes from everyone everywhere all at once
In that order

I would have protected her from getting bullied! I would have made a whole shelf (by hand I like woodwork) to put her plushies on. We would be a deepseek only household...

>>106230523
>>106230523
>>106230523

>>106229563
If I can't run it on at least an even numbered quant with 128gb ram and a 24gb of vram, I don't care.

>>106227955
https://rentry.org/imagepromptgenerator
Warning that it may generate responses that are longer than you are willing to wait for.

You know what...

Fuck it.

Stay tuned.

So what's next for "Open" "AI"?