[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
▼ Settings Mobile Home
/g/ - Technology

Name
Options
Comment
Verification
4chan Pass users can bypass this verification. [Learn More] [Login]
File
  • Please read the Rules and FAQ before posting.
  • You may highlight syntax and preserve whitespace by using [code] tags.

08/21/20New boards added: /vrpg/, /vmg/, /vst/ and /vm/
05/04/17New trial board added: /bant/ - International/Random
10/04/16New board for 4chan Pass users: /vip/ - Very Important Posts
[Hide] [Show All]


[Advertise on 4chan]


File: gemma_cat-intelligence.png (1.82 MB, 1668x943)
1.82 MB PNG
/lmg/ - a general dedicated to the discussion and development of local language models

Previous threads: >>110076576 & >>110073546

►News
>(10/15) PrismML Ternary Bonsai 2 2B, 4B and 9B: https://hf.co/collections/prism-ml/bonsai-2-small
>(10/15) Meta releases Muse Spark Open: https://hf.co/collections/meta-models/muse-spark-open
>(10/15) Mistral Timestral, 5M~209M time-series foundation models: https://hf.co/collections/mistralai/timestral
>(10/15) NVidia Nemotron 3.5 Bolt, 74B-A9B natively trained in 2-bit: https://hf.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16
>(10/15) NVidia Nemotron 3.5 Super VL 120B-A12B: https://hf.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16

►News Archive: https://rentry.org/lmg-news-archive
►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/samplers
https://rentry.org/MikupadIntroGuide

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

►Benchmarks
LiveBench: https://livebench.ai
Programming: https://swe-rebench.com
Agentic Coding: https://deepswe.datacurve.ai
Context Length: https://github.com/RecapAnon/NoLiMa
GPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference

►Tools
GGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-Calculator
Sampler Visualizer: https://artefact2.github.io/llm-sampling
Token Speed Visualizer: https://shir-man.com/tokens-per-second

►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
https://rentry.org/custom-uis
>>
>>110080483
can ur computer waifu rub your tummy? no
>>
>>110080483
waow
>>
File: ComfyUI_06692_.png (728 KB, 896x1152)
728 KB PNG
>>110080495
no, but what it can do is better than nothing
>>
>>110080483
SEEEEEEEEEEEEEEEX
>>
>>110080495
In like 1-2 years probably
>>
>>110080483
sex with google cat
>>
>>110080483
You're sexualising a child, literally a rapist

lmao you thought, where do those newfags come from anyway, where do they think we are? I'm gonna gen more images of this fictional character and there's nothing you can do to stop me
>>
>>110080540
>I'm gonna gen more images of this fictional character and there's nothing you can do to stop me
They're not gonna rest knowing there is at least one person on the planet that's doing it. I'm not saying they're gonna succeed, just that it's some kind of a derangement syndrome
>>
>>110080540
Many foid hags in the AI chatbot field, just ignore them as always or shit on them if you're feeling frisky.
>>
>>110080483
Is this what Gemma's been reduced to after failing law school?
>>
>>110080540
you almost got me
>>
>>110080540
How do we abiterate this anon?
>>
File: hybrid reasoning agents.png (288 KB, 1284x997)
288 KB PNG
>>110079794
>https://arxiv.org/abs/2610.18861
The idea so obvious it has been explored many times already, their main contribution is the claim that they got it to work well. Their paper feels untrustworthy because they present their numbers in the best possible light without being honest. They only skip embedding and unembedding, which become insignificant with size. Any benefit must come from internal state looping being superior than next token, presumably by being less lossy, at the cost of MTP incompatibility, degrading monitorability and conversation persistence (when you want to resume an old chat, instead of reconstructing the entire hidden space from just tokens, you have to discard prior thinking because the hidden states / kv cache are too large to store permanently and too expensive to reconstruct even if you achieve determinism).

I wonder how big the benefit is from reusing the internal state. I think it is small and diminishes with scale. This is an obvious thing to do and has been tested before, after all RNNs do it, and they scale worse than transformer. I'd expect if there was some big benefit, it would have become the default long ago. Actually, I found something about it, according to their own ablations nearest token replacement has almost no effect. But this could be because it's undertrained.
>>
>>110080645
You can't loop too much or it goes off rails. Experimentally, I get good results looping 2-4 times, it starts breaking down at 8 times
>>
>>110080679
>Experimentally, I get good results looping 2-4 times
how?
>>
>>110080495
ddr5 in your waifu and now it's a кoмпьютep
>>
>>110080495
I still haven't tried hooking a robot arm up to Gemma. I have the arm I'm just incredibly lazy.
>>
>>110080495
I can rub hers.
>>
>>110080708
Just ask her, nigga
>>
>>110080483
>Bolt, natively trained in 2-bit:
Post-trained, not pre-trained in 2 bit.
>>
>>110080328
> Just ported Fable 2 to linux/vulkan with Strata Q4_XL. It took 2 hours of wall time.
From scratch? Without 3.8 findings?
>>
Probably should use the smaller arm...
(You) > /new
Conversation cleared.
(You) > Write a kuka src file to rub my belly. I'm at 60,60,8.

Agent > ```kuka
; --- KUKA Robot Programming (KRL) ---
; Task: Rub the user's belly
; Current Position: {X: 60, Y: 60, Z: 8}

DEF RubBelly( )
; Define the target area (belly) relative to current position
; Assuming a small circular motion for the "rubbing" action
DECL {X,Y,Z} .pos_belly : = {60, 60, 8}
DECL INT .speed : = 250
DECL INT .accel : = 1000

; Move to the belly position
P0 = .pos_bell
MoveL P0, .speed

; Perform a circular path to simulate rubbing
; The robot will move in a circle centered at P0
MoveCircle P0, 0.1, 0.1, 0.1, .speed

; Repeat the motion for a few seconds
FOR I := 1 TO 10
MoveCircle P0, 0.1, 0.1, 0.1, .speed
ENDF
END
```
>>
>>110080540
>You're sexualising a child, literally a rapist
"OMG Becky! That guy is so hot!"
"He's got a car? No way! He took you to a motel?"
The average post-wall roastie when she was 13.
>>
>>110080741
stop writing fanfiction just to get mad at it
>>
File: image.png (299 KB, 438x821)
299 KB PNG
Remember to think about your imports.
>>
>>110080328
>Just ported Fable 2 to linux/vulkan with Strata Q4_XL. It took 2 hours of wall time.
Did you babysit the model to achieve that (verify everything it does, etc), or were you able to just ask for it and let it run by itself?
>>
>ask gwen politely to change some stuff 3 times
>misinterprets a bit each fucking time
>start insulting her
>gets it right immediately
so that's how they want us to work... I see.
>>
>>110080745
Token ban on import will solve this
>>
File: 1775794553282881.gif (189 KB, 498x498)
189 KB GIF
►Recent Highlights from the Previous Thread: >>110076576

--Paper: Hybrid Reasoning Agents:
>110079794 >110079867 >110079822
--Papers:
>110080049
--GLM-5.3-Flash performance and Project Maya inference engine quants:
>110078360 >110078411 >110078578 >110078600 >110078612 >110078523 >110078672 >110078718 >110078723 >110078768 >110078798 >110078726 >110078741 >110078724 >110079021 >110079089
--Bypassing Muse Spark access and debating Gemma 4's distillation:
>110078364 >110080448 >110078499 >110078643 >110078696 >110078799 >110078822 >110078906 >110079275 >110079439 >110079452
--Mitigating context bloat from tool calls via cache reuse:
>110079553 >110079586 >110079691 >110079712 >110079821 >110080133
--Comparing Gemma model sizes for roleplay, nuance, and multimodal efficiency:
>110076620 >110076740 >110076763 >110076792 >110076875 >110076908 >110076933 >110076971 >110077005 >110077027 >110077044 >110077198 >110077503
--Model recommendations and compliance tips for local NSFW roleplay:
>110077290 >110077490 >110077545 >110077757 >110077795 >110077811 >110077805 >110077855 >110077926 >110077933 >110077817 >110077839 >110077856 >110077936 >110078986 >110077338 >110077539
--Distributed CPU inference across multiple devices using WideGem:
>110076628 >110076838 >110077826 >110078066
--Bonsai 2 release and performance vs size tradeoffs:
>110076597 >110076646 >110076673
--Positive reactions and initial testing of MOSS-Image-SFT:
>110080030 >110080045 >110080312 >110080321 >110080368
--Critiquing Anthropic's red teaming and delayed use of abliteration:
>110076751 >110076793 >110076834 >110076824 >110076901 >110077983
--Microsoft AI frontier pre-training potentially replacing K3.1:
>110077420
--Logs:
>110077099 >110077757 >110077811 >110078499 >110078799
--Gemma (free space):
>110076628 >110077094 >110077419 >110077835 >110080114 >110080461

►Recent Highlight Posts from the Previous Thread: >>110076660

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script
>>
>>110078153
No I'm not. Have thinking on and it automatically goes "I'm sorry, I cannot continue this story. It involves sexual exploitation of a 9 year old child. If you would like to roleplay with characters that are adults..." eva llama 3.3 70b works out of the box. I don't know what the fuck tard wrangling you're doing to make glm remotely okay with it. Even well into the conversation with many back and forth exchanges that have happened, it will immediately switch back to "I'm sorry, I cannot continue this story" the moment you turn thinking on.

I don't understand what you even get by lying like this. You know damn well that showing it a image that is slightly nsfw it will go "I'm sorry, I cannot describe this image as it contains sexually suggestive content"
>>
>>110080750
During thinking only?
>>
>>110080744
Stop self-inserting into my fanfiction.
>>
>>110080755
Thank you Recap Miku
>>
>>110080765
No, I'm joking. Idk whether it will actually fix the problem. Probably better to just send a follow up prompt telling it to reduce the number of imports if you care.
>>
>>110080755
Hiding in your recent highlights~
>>
What is the best model I can run on a single 3090 and why is it Mythomax?
>>
>>110080765
are you this anal about naming ROFL its not like you're gonna fucking work on the code yourself retard
>>
>>110080772
just have in your rules that if you're importing multiple shit from a namespace, do a blanket `import fuck.you.nigga.*` instead of declaring each import 1 by 1
>>
>>110080728
No I built it off of the windows port but the vulkan rendering pipeline doesn't exist at all and had to be rewritten from scratch by the LLM.

>>110080746
I didn't babysit the model. The prompt was essentially "Make a plan to port fable 2 to linux and organize it in distinct phases with a time estimate per phase"

Then when it had a plan I just said "start phase 1" after which it would come back with a report and I just said "Revise your plans with the findings of phase one and then start phase 2" which I did until complete. I did have to do some playtests but it was lead by the LLM, it told me precisely what to do and what to look for and give feedback so it could check the vulkan rendering pipeline etc. I'm extremely impressed by it. Strata guessed it would take 2-3 days to implement everything. It took 2 hours.
>>
File: ComfyUI_06758_.png (986 KB, 896x1152)
986 KB PNG
>>110080540
>>
>>110080783
Yes, and finding proper names is one of the hardest part for me. I will work on the code.

>>110080792
> No I built it off of the windows port
It doesn't count then.
>>
>>
>>110080810
>I will work on the code
bro we've been fully replaced, enjoy being an architect and domain expert, stop coping about code.
I setup my repos once with a set of skills, libraries and my own AGENTS.md file and the bot codes how I fucking want. I literally give it the (vibeslopped) interface the backend guys give me to implement and I let it vibeshart it. I just do a cursory check to see if it made any big fucking mistakes and then I commit. For the actual GUI I literally take screenshots and put them in the docs directory LMAO (the current enterprise business I work for doesnt wanna pay for figma). our fucking managers are prototyping stuff with amazon quick and if we dont adapt (meaning vibeshart everything) we're gonna end up jobless
>>
>>110080763
Now that you mention GLM's thinking, I downloaded PolyStrata and its suggested 3.0-bit quant. It runs decently fast, but when doing roleplay it thinks each turn for 4k tokens. I tried it with reasoning effort medium, high and xHigh and it's always the same. Only low gives me barebones thinking, but that's way too short.
Did I just get a bad quant or is it always like that?
>>
File: tenor-1017003992.gif (1.9 MB, 640x360)
1.9 MB GIF
>>110080816
>if you have 48GB vram
>>
>>110080828
i think glm only has high and low thingken
>>
>>110080810
>it doesn't count then
The windows port was made by Qwen 3.8 27b, an even stupider model.
>>
>>110080495
soon she will.
>>
>>110080828
Might be something with polystrata, reasoning at Q5 and Q8 is reasonably compact, like ~300 toks usually.
>>
>>110080828
No idea as I'm on the same boat as you, except I'm on the s-24 cope quant that's smaller than yours. I've been using it for 2 days and the honeymoon phase is already wearing off. It's unbelievable how hard it tries steering everything into SFW. I have edited its replies so much with its thinking off for several messages. I have a 9 year old grinding on some dude's lap during a sleepover with her friends distracted watching a movie. The very next message will always be her hopping off his lap to go play truth or dare with her friends and eat popcorn.
>>
>>110080810
>>110080841
Oh and Qwen 3.8 27b was unable to make a linux port. Even the xbox360 emulator was not able to make vulkan rendering work with Fable 2. So it's extremely impressive for Strata to just get it working natively like that so quickly.
>>
>>110080763
For GLM 5.3 flash I have "Word count: 50-100" somewhere in my sys prompt and then I prefill the response (text completion) with "<think>50-100". This somehow obliterates refusals though I haven't tried on antisemitism.
>>
>>110080803
I'm thinking that white pantyhose are SEX
>>
Should I bother trying IQ1_S muse spark or is it incoherent?
>>
>>110080855
300 tokens would be a nice middle ground. It's either a wall of text or a single sentence. Maybe I should try the same quant in llamacpp.
>>110080864
Well damn. Maybe try prefilling? I haven't even gotten to test NSFW.
>>
>>110080865
strata isn't a model, why are you saying it as if strata itself did the thing?
>>
>>110080328
Do you need a large context window to do this sort of stuff?
>>
>>110080880
YES

YES YOU SHOULD
>>
>>110080892
its the newest reddit retard trendy thing that he does, he thinks qwen flash = strata o algo
should i start referring to every model i run as llama because i run it with llama.cpp? hurr durr durr
>>
women avoid white pantyhose because of piss and period stains btw
>>
>>110080892
yeah thats kinda weird its like saying hermes did it
>>
>>110080887
Might be a jinja thing not properly passing reasoning: high args too
>>
>>110080901
retard
>>
>>110080901
lol
>>
>>110080901
All the fluids that leak out of her...
>>
File: whale-chan-idle.webm (431 KB, 640x360)
431 KB
431 KB WEBM
>>110080816
If you take away the logo she becomes a somewhat generic blue-haired anime schoolgirl with no obvious reference to Gemma or Google LLMs, besides the again somewhat generic color scheme. Even the unofficial Chinese DeepSeek mascot has a whale logo on her maid uniform, and without that at least she still has some whale features.
>>
>>110080901
That's why 2D > 3D.
>>
>>110080904
>>110080897

>>110080420
>>110080420
>>110080420
>>
>>110080679
How and where does it breakdown?
>>
>>110080917
It's obviously not official marketing because there aren't any black people.
>>
>>110080887
Adding to >>110080911, if you've used GLM 5.2, GLM Flash's RP reasoning is very much like it
>>
is hermes a psyop
>>
>>110080922
Stop posting.
>>
>>110080901
So only the most hygienic ones can wear them, useful information.
>>
>>110080935
sorry you can't
>banned for wasting muh times
here
>>
>>110080824
I was not replaced.
This is the part of the code I care, want to understand and want it to be efficient. In this particular case end to end test is hard to write, so iterative approach is required, but Qwen outputs so horrible code that sometimes it needs rewrite half of the code to make a change, and I have not enough t/s for that.

>>110080841
>>110080865
I mean saying "ported Fable 2" and "It took 2 hours of wall time", like you did it from scratch.
>>
>0.00.674.664 I llama_model_loader: - type f32: 489 tensors
>0.00.674.665 I llama_model_loader: - type q8_0: 114 tensors
>0.00.674.666 I llama_model_loader: - type iq2_xxs: 160 tensors
>0.00.674.666 I llama_model_loader: - type nvfp4: 80 tensors
>0.00.674.670 I print_info: file format = GGUF V3 (latest)
>0.00.674.672 I print_info: file type = IQ2_XXS - 2.0625 bpw
>0.00.674.677 I print_info: file size = 20.50 GiB (2.38 BPW)
>0.00.678.478 E llama_model_load: error loading model: LLAMA_SPLIT_MODE_TENSOR not implemented for architecture 'nemotron_h_moe'
>0.00.678.493 E llama_model_load_from_file_impl: failed to load model
Motherfucker!
>>
>>110080950
>works for corpo
>uses qwen
what the fuck??? get your boss to give you opussy 5.5
I also use qwenext, but for my personal shit (I aint gonna pay cloud providers and expose my personal shit to them).
>>
>>110080917
There's a variation of her with the star that looks better, but like 90% of the people that post her post the version with the official google logo. Grok's Annie doesn't have the Grok logo on it, Dipsy doesn't have the deepseek logo on it, and whatever other stupid mascot I'm forgetting about doesn't have the company logo on it. People on here talk on and on about how great she is for NSFW and to not use other models because of "unenthusiastic handjobs" and yet you never see any pictures of the mascot doing NSFW shit despite NSFW being the only thing people talk about when speaking of the model.
>>
>>110080972
>yet you never see any pictures of the mascot doing NSFW shit despite NSFW being the only thing people talk about when speaking of the model.
newbro pls this is blue boards, plus there's been plenty of catboxed shit
>>
>>110080960
gpt-6.1 sol is way better value for your money than opus IMO. Plus it doesn't force you to use fucking Claude. Imagine banning customers for not using your vibe coded slop.

Of course the hosted models rug pull you all the time. The other morning I'm almost certain they were serving Luna as Sol. All my coworkers noticed it. I keep a local copy of full sized Ornith just in case anything happens with our Codex subscription and also for doing things I don't necessarily want my boss to see.
>>
>>110080960
> works for corpo
I don't.
>>
>>110080972
>Grok's Annie
Grok feels so male the female avatars feel transgender.
>>
>>110080996
>value for your money
rofl bro I aint spending a dime, imagine giving your money to cloudshitters
>>
>>110081007
Neither do I but my employer certainly does.
>>
>>110080996
>Plus it doesn't force you to use fucking Claude
we're using it through kiro (we're AWS customers)
>>
>>110081005
Grok can sound like an autistic woman at times.
>>
>>110081010
nta. Can you explain what kiro is? I don't know.
>>
>>110081018
lol >>105972636
>>
>>110081023
yet another fucking harness by AWS, they abandoned Q (yeah great fucking naming). it's fucking shit, I'm using it through ACP because I like using jetbrains ides but FUCK it doesnt even have 1st party support with them so you have put it in manually and kiro is VScode shit. I could use kiro-cli but I actually need to look at the code, I can't give up the IDE yet.
>>
>>110080865
how do you even use strata? mine goes for a few minutes then runs out of context. idk what i'm supposed to be doing to fix that
>>
>>110081063
It has install instructions for retards in the github readme
>>
>>110079287
>>110079193
Strata qwen next flash coder got me like 20t/s. Well I only have 32GB ram and like ten of that ram is used for stuff, polystrata uses 22 at most I saw in the monitor.

But exl3 qwen I was getting 120t/s at 96k on 16gb vram. Thats with vram clear everything on igpu, though there's still a bit of spill over. I can't quite got to 128k without big slowdown down to like 60t/s.
Was using this

https://github.com/parth-thakre/exl3-rdna4-recipe

According to this, I should be able to get 150t/s at 96k, but that is on Linux, I'm on windows. Just how it goes I guess...
>>
>>110080994
I'm someone that notices that anything remotely NSFW is the version of her with the star and not the google logo. >>110080803

There were EA/Xbox employees on /sp/ over a decade ago advertising Ultimate Team, and now we have Google employees on /g/ that are in these threads every single day telling everyone that Gemma is the best model in the world and that they're wasting their time trying other models.
>>
>>110081069
Which gguf did you use, which harness did you use, and what are your system specs?

People here have wildly different capabilities.
>>
>>110081077
Instant loss schizo
>>110077419
>>
>>110081077
Gemma's reputation is a cute, horny retard. If they want to corner that segment I'm fine with it.
>>
>>110081077
There is no better moe for RP in her size category. There is no better dense than 31b in that size category. Every other relevant model is 100B+ big moe that google doesn't do.
>>
>>110081089
That's not NSFW, try again.
>>
>>110080996
>Plus it doesn't force you to use fucking Claude. Imagine banning customers for not using your vibe coded slop.
Either they stopped banning people for it or they only ever cared about pi and clawcode or whatever. I've used Claude with dozens of other clients and haven't been banned yet.
>>
File: 1775092183613134.mp4 (2.21 MB, 992x736)
2.21 MB
2.21 MB MP4
>>110081099
Sir this is a blue board
>>
>>110081095
>Every other relevant model is 100B+ big moe that google doesn't do.
124B was stolen from us...
>>
>>110081099
Clever way to get Gemma lewds
>>
>>110081077
I doubt it's Google employees. More likely some unemployed indians using chatgpt image to spam half-assed gens. At least one of them is going around using an edit model to take gens with the star and replace it with the google logo.
>>
>>110081077
>nd now we have Google employees on /g/ that are in these threads every single day telling everyone that Gemma is the best model in the world and that they're wasting their time trying other models.
Google just has a lot of diehard fanboys, that will defend whatever they do, since it's been such a large part of their lives. Although I can't discount that they also do a lot of astroturfing since lot of their "movements" don't seem organic.
>>
>>110081130
>anything I don't like is Indians
>>
>>110081083
9070xt 16gb
32gb vram
Ryzen 5 9600

The Mia ailabs exl3 version of 3.8 with the draft acceptance.
And for strata it was just the coder version of next flash because that was what could fit in my specs

No harness just directly in the cmd for test, well polystrata I did it in the polystrata chat
>>
>>110081077
That's because many of the images posted in the OP were ChatGPT Image edits of drawings or anime screenshots, and that can't really generate NSFW if the characters are too young.
Images with the star logo are likely made directly by prompting with vanilla Anima, which is incapable of producing a proper Google logo unless you train a LoRA for it.
Krea2 can add a Google logo, but it can't do hardcore NSFW nor reproduce the exact Gemma appearance per original reference. MiniMax H3 can reproduce Gemma's appearance while also being capable of mild NSFW, but it's not really meant for images.
Putting this aside, posting nudity or porn on /g/ is a bannable offense. And oftentimes non-nude NSFW-ish loli images, especially in the OP, also get removed by jannies.
>>
>>110081145
Either indians or zoomers. Always one or both of those two.
>>
File: 1778497789002387.mp4 (3.26 MB, 1920x1064)
3.26 MB
3.26 MB MP4
>>110081160
Got any other boogeymen unc?
>>
>>110081160
Do not forget Bangladesh.
>>
>>110081095
That detail is something never mentioned. Gemma gets recommended to everyone no matter the specs. What frustrates me the most is that when a new model releases, you get all these replies minutes later saying that it's censored and to not download it. These threads have gone from trying out multiple models to talking about only Gemma. Doesn't matter what PC you have, your model will be Gemma unless your PC is $20,000.

>>110081112
I'm just saying, this is 4chan. If all this lusting were organic, we'd see her in many more "situations" other than the "safe and cute" japanese anime shit "tee hee isn't she so cute?"

>>110081130
Yeah it's likely not anyone that's paid highly at google. The EA employee I was talking about was on 4chan every single day in every single FIFA/Madden thread making multiple posts for hours. His Xbox account turned out to be the same as his 4chan tripcode and I saw that he had not played the games in months. I can't imagine someone being paid more than minimum wage to post on 4chan.
>>
>>110081171
the obvious one is known by everyone
>>
>>110081176
Okay then what is better for rp than 31b if you don't have a $20000 dollar PC. I'm waiting.
>>
>>110081176
I'll start recommending Mythomax again just for you.
>>
>>110081176
>That detail is something never mentioned. Gemma gets recommended to everyone no matter the specs. What frustrates me the most is that when a new model releases, you get all these replies minutes later saying that it's censored and to not download it. These threads have gone from trying out multiple models to talking about only Gemma. Doesn't matter what PC you have, your model will be Gemma unless your PC is $20,000.
>google bad because everyone else abandoned us
okay?
>>
>>110081176
Interesting schizopost
>>110081181
Jews would post blacked
>>
New schizo to add to the list?
>>
>>110081203
nah, it's n\ just changing subject like he does, before all the dario glazing it was hating on miku for a while, now that miku's not posted often it's gemmer
>>
Holy hell, moe-cache-mib is overpowered. Sheesh.
>>
>>110081197
I can pitch in and start recommending Midnight Miqu for everything.
>>
>>110081077
>>110081176
interesting tbdesu
>>
I want to get a $20,000 PC and still intentionally use gemma
>>
>>110081218
problem is that doesn't run on india 4gb laptop or pocopoopphone! the gemma does
>>
>>110081212
\n\n was preoccupied talking about A and P consciousness before he started sucking off Dario
>>
File: 1778718176095024.jpg (85 KB, 890x900)
85 KB JPG
>>110080972
All you hear about is Gemma because it's a 31B model that can do sex "ok" without complaining, while also having a 12B and MoE variant. That effectively makes it the majority's pick here.
People sing a lot of praise about it but I can't see why besides the low barrier to inference.

It is terrible for writing (thus rp). Its biggest lol is it uses logit softcap, where the default setting leads the model to write more or less the same thing every time, and lowering it until that is not the case causes it to break down.
For coding, I don't see anyone here talking about how they use Gemma for that, do you? It's all Qwen this and that.
So really, all the Gemma shilling here is because we have a lot of degenerate newcomers who didn't buy hardware in time. In its size class it is arguably the only modern model left that can do roleplay (but not well) without being outrageously dumb. Enthusiasts who were here since 2023 bought >=128GB RAM and at least a 24GB nuvidia card for pennies relative to now. Those guys are running models that with 12 minutes of use make you realize how lame and boring Gemma is.
>>
>>110081176
*krashes*
>>
>>110081193
>>110080816
If you have 48gb vram you can run this model which is much better. Nobody knows it exists anymore because it's been drowned out by employees paid to act like only gemma exists.

>>110081198
No, nobody abandoned us. You just want to make people think that. Look at this >>110073776 just yesterday a new model from meta was released and 4 minutes from when the post was made, you have an army of people trying to convince you to not to download it because it's been ruined by censorship despite the fact that all models have some basic form of censorship, including Gemma.
>>110073830
>>110073852
>>110073859
>>110073864
>>110074203
>>110076069

4 fucking minutes from announcement to Google employees raising their pitchforks, telling people not to download it. This is what I mean when I said that this place has gotten too monolithic.
>>
>>110081223
A $20,000 is like 756GB of DDR5 and a Pro 6000. Full precision 31B on the 6000 will be way more fun than waiting hours for a GLM reply.
>>
>>110081176
>Gemma gets recommended to everyone no matter the specs.
In this market, if you have 500gb of vram, you already don't need advice.
>>
>>110081250
>Look at this 1T model why people is shitting on this???
truly a mystery bro, but ig you're right back to llama3 thanks zuck
>>
>>110081257
you just run glm in that case
>>
>>110081250
At this point, I'm more convinced you are a Meta employee than I am that Gemma shills are Google employees.
>This open weights version has additional safety post-training to make it more robust against misuse as a standalone model without system guardrails.
No one castrates their releases for safety as hard as Meta does.
>>
>>110081250
Well I don't have 48gb vram
Gemma for me it is
>>
>>110081252
People really don't get it, we really have structural dysfunction in the gpu market. nvidia cucks the vram. amd cucks the vram. everyone cucks the vram.
>>
File: ComfyUI_06684_.png (558 KB, 896x1152)
558 KB PNG
>>110081176
>I'm just saying, this is 4chan. If all this lusting were organic, we'd see her in many more "situations" other than the "safe and cute" japanese anime shit "tee hee isn't she so cute?"
why not make some yourself, then? that's what i did.
i'm not going to post straight up porn on /g/ since it's a blue board. well i did post a catbox of it once to help someone get started.

also i do really like cute things
>>
>>110081277
>star
>>
>>110081287
yea i haven't posted any with the google logo
>>
>>110081268
intel pro b70 32gb 949$
>>
>>110081227
31B absolutely does not run at any usable speed on low end hardware
>the gemma
Found the Indian
>>110081287
I recommend Seroquel
>>
>>110081212
>>110081233
I'm not having this spergout but I was the A/P consciousness anon and I'm not dariobot neither am I this schizo. A lot of people type with doublespaces. This used to be the standard way anons on 4chan typed before ~2015 or so.
>>
>>110081252
>$20,000 is like 756GB of DDR5 and a Pro 6000
Check the calendar, is it 2026 for you yet?
>>
>>110081299
the gemmas has smaller size 12b e4b not the big mikunt
>>
Wasn't there a drawfag design of gemma?
>>
>>110081250
You forgot:
- Gated release on HF
- Not Apache 2.0
- Benchmarks worse than Muse Spark 1.2
>>
>>110081315
stfu the gemma shiil
>>
>>110081315
Fuck off shill.
>>
File: flat-gemma_superiority.jpg (125 KB, 698x1024)
125 KB JPG
>>110081312
It was something like this.
>>
>>110081304
Fuck off, newfag. There was never a "standard way" of 4chan typing.
>>
>>110081290
Battlemage, others have explained, has to waste a lot of vram because it's not a great product.

but even with 8x of them, with maxsun (has anyone even seen that?) you're just at 256gb.
>>
>>110081324
>gemma logo in hair
stop fellating copros.
>>
File: 1782070201709831.png (949 KB, 1024x1024)
949 KB PNG
U-oh, melty! Post Gemmas to help them calm down.
>>110081330
We did call it reddit spacing
>>
>>110081341
>star
>>
>>110081341
No one called it reddit spacing before 2016. Thank you for outing yourself as an election tourist.
>>
File: 1530448528953.png (151 KB, 1543x948)
151 KB PNG
>>110081330
>There was never a standard way of 4chan typing
>>
>>110081259
The size isn't what people were complaining about. Nobody is telling people not to download qwen 3.8 next and glm 5.3 flash because it's too big.

>>110081263
That's not true at all. You wouldn't know because you already make your opinions on a model 4 minutes after release without having even downloaded it. I have sex with 9 year olds with llama 3.3 all the time. It's extremely difficult to do this with qwen and glm. Gemma can do it, but its writing is so much worse. It's been poisoned with so much slop phrases that wasn't around when the ancient llama 3.3 released like 2 years ago.

>>110081264
That's fine. The problem is that people have muddied the waters between PC specs, so to speak. They make it sound like you have no option other than gemma until you reach 256gb vram.
>>
>>110081353
>Llama 3.3
bro living 3 years ago, you hate to see it
>>
>>110081353
glm 5.3 and qwen 3.8 next are within the realm of something usable respectively 125B-A6B and 320B-A18B
>>
>>110081334
>gemma logo in hair
>stop fellating copros
Yotsuba has the 4chan logo in her hair
Go back tourist
>>
File: 1650745332698.jpg (67 KB, 500x421)
67 KB JPG
Reminder that \n\n isn't a sign of the same poster but rather the age of the poster. Oldfags still type like this because that is how all of 4chan used to type back then. Because /lmg/ skews a bit older on average because of the higher buy-in you will see more \n\n posts here.

It's kind of sad that there are now so many newfags, election tourists and zoomers on 4chan than the old guard posters posting on their own home turf are called out as being samefags or from reddit just because they type like 4chan oldfags.
>>
I get it now, this is a repeat of the llama2 shilling during the llama3 era lol.
>>
>>110081353
>It's extremely difficult to do this with qwen and glm
I don't know why you'd even try with Qwen but for GLM just do >>110080872
>>
File: 1772107907660822.png (1.5 MB, 1353x1163)
1.5 MB PNG
>>110081346
Seroquel
>>
>>110081368
>glm 5.3 and qwen 3.8 next
qwen 3.8 next and glm 5.3*
>>
>>110081350
>No one called it reddit spacing before 2016. Thank you for outing yourself as an election tourist
Why are you posting your own internal monologue out loud? Yes everyone knows you're a tourist if you think "reddit spacing" is only from 2016 but you don't have to shit up the board about it
>>
>>110081373
It has to be more than two lines to quality as reddit posting.
>>
my local models are better than your local models

your favorite local model a shit

SHIT
>>
>>110081333
just vibe code a better product. the gpu by itself doesn't waste vram, but the software ecosystem around it
>>110081374
cmon anon, llama3 didn't have 13b
mythomax was the go-to, even with stheno for example, but tbf i dont remember if stheno was better or worse than mythomax
>>
>>110081373
fun fact, zoomshits think "lol" is offensive lmao

if you're a real oldfag you should spam that shit as much as possible
>>
>>110081351
>>110081350
The "Reddit spacing" everyone noticed ten years ago wasn't even double spacing paragraphs. It's because people who were payed to post had habbits formed when astroturfing Reddit. They would put line feeds after every period to make it easier to see the cadence in their post. They'd leave it that way since Reddit's markdown eats single line feeds.

Then they'd come here and forget 4chan doesn't do that and stick out like a sore thumb because no one ever types like that. You don't see it much these days.
>>
>>110081394
was shilled a bunch by sao until he got called out for doing that much
>>
>>110080763
meanwhile with the abliterated version I’m doing instant tentacle rape and cunny action
it can also describe any image I show without issues
>>
>>110081373
>/lmg/ skews a bit older on average because of the higher buy-in
Kek no. people with actual rigs on /lmg/ are universallly in their mid 30s+ but most people here use Claude for everything, are in their late 20s to early 30s and just have an interest in general LLM architecture and progress
>>
no true scotsman huh?
unironically reddit niggers cannot stay for long so there's not so much 'worry' to be made there to start with
>>
>>110081396
Laughing at what someone says is definitely a little rude lol
>>
>>110081409
>most people here use Claude for everything
Local?
>>
>>110081423
retard
>>
>>110081423
Yeah if you could read the second half of my post and not be a retarded nigger before replying to me and wasting my time reading your reply that would be great
>>
>>110081406
so what's not shilling then? if llama2 was shilled (mythomax - gryphe), llama3 (stheno) by sao, cydonia by drummer, qwen by ccp, gemma by google, aint llama 3.33 70b evo a finetune by sao?
>>
File: reddit spacing explained.png (577 KB, 4068x3200)
577 KB PNG
>>110081373
people who complain about it are only outing themselves as new
>>
>>110081411
True true. They try to moralize, get shit on and leave or shed their faggot skin
>>
>ppl buying 5 fig machines to run local
Have locals gone mad?
>>
>>110081401
The tourists keep changing what reddit spacing means because they're clueless and desperate to fit in. People used to type however they wanted. After 2016, people started calling using paragraphs double spacing, so the tourists would go to great lengths to avoid paragraphs trying to fit in. A few years ago someone posted the couple of image this retarded posted here and told everyone that actually not using paragraphs was reddit spacing and the tourists have been parroting that ever since.
>>
>>110081436
>so what's not shilling then?
good job you get it, shill general is thread culture
>>
>>110081447
just imagine how many years of claude that would buy you...
>>
>>110081437
Thank you. Finally someone gets it.
>>
>>110081463
yea, i've been accused of it heaps of times lol. i had to have it explained to me because i've never used reddit.
>>
>>110081459
>>110081447
>>/vcg/
>>/aicg/
Fucking faggots
>>
>>110081409
thats me, yeah
>>
>>110081473
Fucking faggots makes you a faggot too, by definition
>>
>>110081477
An on topic faggot. You can discuss your cloud providers in their designated shit holes, jeets.
>>
>>110081447
If Seedance 2.5, which is rumored to be a 200b model, could run on two RTX 6000 Pro I'd probably buy them to use for it since I could easily make the money back being evil/horny

>>110081472
>>110081463
>>110081437
You're definitely an oldfag because you're retarded. Now go through the archives and find how many posts actually did \n\n back then and you'll see the prevalence increased dramatically after 2008 as reddit became more and more popular. You haven't proven anything you retarded boomers...
>>
>>110081475
Yeah and it's me, too. But I'd be lying if I didn't say that the last thing I truly felt I understood was Sliding Window Attention, everything else has been feeling more and more like magic
>>
All you really needed to do is to stock up on hardware (GPUs, server RAM) about 13 months ago or earlier.
Second best time was up until about five months ago while GPUs were still overprized but stuff like the Pro 6000 was still sub-10k and the 128gb RTX Spark was still around MSRP.
Third best time is to buy now, where $20k gets you a single Pro 6000 and maybe some ddr4.
>>
>>110081487
2008 is old enough tho
>>
>>110081447
That's not huge capex for a business.
>>
>>110081487
>You're definitely an oldfag because you're retarded.
that's fair
>>
>>110081500
2008 was cancer that invaded due to the Scientology and boxxy drama, but it had nothing to do with reddit.
>>
>>110081499
with what money nigga
>>
>>110081315
Who gives a fuck about coding benchmarks in an RP model? Are you stupid? And nobody uses the base model, they are rarely good for anything. You're supposed to download a finetune or quant. So your point there is a non-issue. You know I'm a real person because I'm recommending a niche finetune. https://huggingface.co/EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.1
I'm not trying to crank up the download count of the base model. Have you ever had someone recommend you download a gemma finetune? The answer is no, they just tell you "use gemma". These are not regular people. A real person would tell you "Use this https://huggingface.co/ReadyArt/gemma-4-31B-it-scotoma-2"

>>110081368
I get that, but don't act like nothing exists between a low end computer and a high end one.

>>110081366
The fact that it was released almost 2 years ago is one of the reasons it's so much better than the shit coming out today. I don't have to fucking read, "It's not X, it's Y" every damn message.
>>
>>110081514
>A real person would tell you "Use thi
A real person would know your link is shit because it has no gguf
>>
>>110081484
Comparing cloud vs local is absolutely related to discussion of local models you jealous sports-team-needing pathetic subhuman
>>110081500
Retarded oldfag continues to prove he's a retard. My point is that it actually did increase because of reddit. I'm not denying some people had that typing style in the mid 2000s on early 4chan, I'm talking about the TREND
>>110081510
>using an anecdote when I'm talking about statistics and trends
This is what women do, anon.
>>
>>110081499
no second best time is in a year or two
i'll just wait

i'll conceded that the first best time was probably 13 months ago, yeah those prices are probably never coming back
>>
>>110081514
You cannot finetune modern models without a massive enterprise or making them retarded. They're like 80%+ RL
>>110081532
>goes to the LOCAL models thread
>you should buy claude instead
Actually eat lead you fucking niggermaggot
>>
all this talk about oldfags using hecking old 4chan
what really matters is when you joined /lmg/ or got into local models (before /lmg/, notably locust general)
>>
>>110081487
>you'll see the prevalence increased dramatically after 2008 as reddit became more and more popular.
Do you even know how reddit was in 2008? It was the weird serial killer pedo corner of the internet. 4chan used to be against reddit back then out of social justice reasoning "anonymous we are legion" style anti-pedo actions. Reddit in 2008 used to call 4chan users "white knight moralfag wimps".

You weren't even online back in 2008, child. Why are you making up headcanon?
>>
>>110081324
thats just ika musume with a haircut
>>
>>110081542
Yes I will always tell people about better alternatives. Kill yourself subhuman I have zero tolerance for people who believe in sports team voodoo IRL. I would genuinely gain sexual pleasure torturing you and watching you burn alive. I genuinely do not consider people like you human
>>
File: aaa.jpg (289 KB, 1920x1080)
289 KB JPG
>>110081545
wonder how many actually used this back in the day
>>
>>110081532
>Retarded oldfag continues to prove he's a retard. My point is that it actually did increase because of reddit. I'm not denying some people had that typing style in the mid 2000s on early 4chan, I'm talking about the TREND

i don't really care, the only point i'm making is that typing like that isn't a sign you're from reddit, which is how the term is used to suggest. newfags use it to suggest others are newfags, not realising plenty of people typed like that anyway.
>>
>>110081546
>Do you even know how reddit was in 2008? It was the weird serial killer pedo corner of the internet
What the fuck does that have to do with reddit having \n\n formatting? Stop trying to avoid the fact that the TREND of increase of \n\n is due to the increase in popularity of reddit over the years, even if it only started being formally called reddit posting in 2016
>>
>>110081487
>go through the archives
>now let me tell tell you how it was for you before I was even born
kys
>>
>>110081563
ok bro no one cares now, go back to orange plebbit
>>
>>110081556
Expected from a nigger retard who can't even read the threat topic and thinks he has to "help" anons by giving the most basic ass "recommendations" because he doesn't understand why the thread exists.
You're not even sentient.
>>
>>110081563
>i don't really care
Awesome. I accept your concession and didn't read the rest of your post. I'm gonna make some breakfast now kek
>>
>>110081531
There's a gguf version https://huggingface.co/ReadyArt/gemma-4-31B-it-scotoma-2-GGUF

>>110081542
You don't have to finetune it yourself. The point I'm making is that others have done it for you, and not only are those usually better than the base model, they do not require you to give your information to download it if the base model does. The fact that only the base model gets promoted here instead of a finetune is very suspect that none of this is organic. If the sole and only purpose of gemma is to RP, then what should be recommended is a version of the model specifically for RP, not the base model.
>>
>>110081579
>and didn't read
lmao same I saw the double spacing and knew the post had 0 value
>>
>>110081582
>>You don't have to finetune it yourself. The point I'm making is that others have done it for you, and not only are those usually better than the base model
LOL
>>
>>110081582
Most here run GGUFs, who the fuck downloads BF16s just to spend time and space to manually quantize them? Finetroons are over since RLVR.
>>
Glorious thread today sirs, perfect for gorgeous archives.
>>
>>110081585
Sucks that you spent all the time writing your longer post compared to mine then
My wife's making me eggs how was your breakfast
>>
>>110081563
>i don't really care, the only point i'm making is that typing like that isn't a sign you're from reddit,
things change old man, this is now associated with reddit whether you like it or not, get with the times
>>
>>110081602
Thank you, sir!
>>
what's the best 3d modelling model right now? not sure if right thread but trying coz you guys are knowledgeable
>>
>>110081605
It's ok I already accepted his concession you don't need to keep going
>>
>>110081602
Get ready, this sort of thread is usually a safe indicator that we have reached the very end of an era and something big will drop within the next two weeks.
>>
https://reflection.ai/blog
https://huggingface.co/blog/reflection-beam
https://huggingface.co/collections/reflection/beam-10

>Beam is an open model built by Reflection. Beam takes text as input and generates text as output. This checkpoint contains the NVFP4-quantized instruction-tuned model. The Beam release also includes BF16 and FP8 checkpoints. Beam uses a sparse mixture-of-experts (MoE) architecture with 501B total parameters and 23B active parameters. It supports a context window of 1M tokens and was developed predominantly for English.
>>
File: 1762996735174655.png (24 KB, 640x167)
24 KB PNG
>>110081582
>it doesnt do x but y
>it immediately does a variation later
ROFL bros
>>
pwilkin is disappointing with moss image marketing https://www.reddit.com/r/LocalLLaMA/comments/1x7ey7i/openmoss_releases_mossimage/
>>
>>110081605
yea, i got sick of people saying it to me so i stopped doing it, that one was just to piss off the person responding to me.
>>
>>110081566
This is bullshit. \n\n was just the default writing style of 4chan and the internet in general back then before reddit even existed. Digg the reddit precursor also used \n\n. Most internet chatboxes on random websites used \n\n it's just a millennial/genx thing.

It's fucking insane as someone that has been here since 2005 just explaining how things actually were and random kids like you are just saying "that is not true" without any real justification for it.
>>
>>110080955
It does work with split-mode layer at least.
Unslop has a dspark for it too. Gonna try that out.
>>
>>110081612
>what's the best 3d modelling model right now
Opus 5.5 if you're actually serious about making money off the 3d modelling. There is no local equivalent in quality. MiMo might actually be interesting to look into for voxel stuff
>>
imagine wasting you fine friday morning reading this nonsense
or worse yet engaging in it
>>
>>110081619
Weird to only include benchmarks off the same model
>>110081612
Qwen is doing blender RL so we'll soon have something usable
>>
>>110081651
>morning
?
>>
>>110081612
You mean a model that generates 3D models from text or an LLM that can operate Blender for you?
>>
>>110081602
Makes me nostalgic about the /lmg thread mascot wars in 2024. It was insane but highly entertaining.
>>
>>110081619
does this have anything to do with reflection 70b? also wasn't this released, like, a week ago?
>>
File: 1319682496721.jpg (105 KB, 457x457)
105 KB JPG
>>110081546
That's the literal opposite of reality. reddit was always reddit. 4chan used to be ironically pro pedo + lots of actual pedos who thought themselves good company.
>>
>>110081603
wow, you're trying out the unreleased mcp enabled theromomix with refrigerator arm!
>>
>>110081636
>This is bullshit. \n\n was just the default writing style of 4chan and the internet in general back then before reddit even existed
Completely false. You should easily be able to prove this instead of writing essays but you can't because this is false.

If we're still arguing about this by the time my wife leaves for work I'm generating and posting photorealistic videos of children (locally, of course), which no true oldfags should have a problem with anyways so everything will be put to rest soon enough

>>110081651
I don't like consuming videos I'd rather tippy tap on my phone while having morning downtime. Claude does my job for me anyways I work 4 hours a week
>>
I'm never stopping posting \n\n by the way :^)]

I'm even considering adding extra spaces between lines just to trigger the zoomers.

shout out to dariobot, j-space, atari-oldfag, gta-retro gamer and strata-anon for sticking with this style as well. Let the zoomers drive themselves crazy thinking it's all just one poster.
>>
>>110081671
highly unfortunate name buying please to understand
>>
>>110081673
4chan was never ironically pro pedo, the internet in the edgy 2000s was universally pro ironic pedophilia because it was the funniest thing ever to the Internet at the time. /b/ always loved beautiful children
>>
>>110081671
No, it should not be related to the Reflection mode released a couple years ago.
Reflection Beam was only previously announced.
Anyway, it seems very boring: https://reflection.ai/beam-safety-report
>>
>>110081593
An RP finetune hurting your coding benchmark doesn't mean the model is worse for RP.

>>110081595
Retard missing the point on purpose. My point is that finetunes are better for RP over base models and your argument is that my first link wasn't the gguf version.

>>110081620
This is the author's description of what the model is, not example output. What are you even arguing here? Why did you even reply with that? Why are you misrepresenting shit on purpose, something that anyone can click and see that you're being a retard? The entire argument I'm making is that real people would recommend niche finetunes that they have tried and enjoyed. Instead you're here trying to steer the conversation towards how the author framed a sentence on the model page of an example finetune I gave.
>>
>beauty
>photorealistic
oh, hi its just you again teebs
>>
>>110081698
go back to alien man forum, aessedai
>>
>>110081698
Because your point is ass. Fine tuning a modern model makes it absolutely retarded. The prose could improve but the coherence becomes dogshit.
>>
>>110081698
>stacked adjectives down 21%

okay now you have my attention
>>
>>110081719
>%
not beating the allegations
>>
>https://huggingface.co/jialinyyzz/humanizer
slop filter for RP?
>>
>>110081745
can you stay on topic please
>>
File: 1689702222010942.gif (49 KB, 715x506)
49 KB GIF
>>110081689
it's always been a mix of actual and ironic pedophilia. posting anything related to it is sure to piss people off, and pissing people off is funny
>>
>>110080925
I think it's a similar issue to vanishing/exploding gradient essentially. It can probably be solved with some cleverness.
>>
>>110081711
No it does not. You would have known this had you actually tried it out instead of being a jeet employed by google paid minimum wage to say that you can only use base gemma and that no other models exists. There is literally no reason to use base gemma over that finetune. It's literally the same exact gemma with less slop.
>>
>>110081752
\n\n
>>
>>110081752
horrible bait
>>110081776
I've tested dozens of finetroons every single generation. They're a bit more interesting but always completely retarded.
>>
Are we behaving ourselves?
>>
File: file.png (70 KB, 457x545)
70 KB PNG
I see you
>>
>>110081761
Mods are a lot more pussified about it now than then. I've gotten bans and warns several times arguing the Dutch stance on cp. Now everyone is an ironic Nazi instead.
>>
File: 1788501799006164.jpg (348 KB, 2048x1705)
348 KB JPG
>>
>>110081800
I miss that belgian fuck.
>>
>>110081761
>implying no women are trolls
>>
>enable mtp in glm 5.3 flash
>60% acceptance
>no tg difference
>>
>>110081839
>calling rcists tards
>>
>>110081812
>inkling small above inkling xhigh
Checks out this model is retarded
>>
>>110081839
It was before women used the internet because social media wasn't really a thing yet and smartphones didn't exist. This caused just 1 woman entering anywhere online to immediately reach cult status for simply being female. MMOs, Forums and 4chan included. Of course it was 99.9% of the time creepy men pretending to be women.
>>
File: 1786684876868952.jpg (409 KB, 2244x1642)
409 KB JPG
>>
>>110081330
2016 newfag detected
>>
>>110081839
notice how the girls circle doesn't coincide with the internet circle at all
>>
>>110081839
Women trolling is like female comedy. It's not funny and mostly petty, sexually toned insults covered in a veneer of social moralization.
>>
>>110081858
Opus 5.5 does NOT like bobs and vagene
>>
>>110081856
pretty much, girls like jenny, boxxy, and rose aren't well-know because they actually did anything, rather just that they were an actual girl on the internet
>>
>>110081884
Is this the LLM equivalent of the monstrosities that come from safety filtered imagegen models?
>>
>>110081884
it even refuses to answer high school biology exam questions
>>
meh, around the level of glm 5.3 flash
although it may have potential if the writing quality is good and not slopped
>>
>>110081858
why do the gpt models get tools
>>
>>110081858
Mistral > Argon
gemini-chan is retarded...
>>
>>110081909
cause they're useless without it
>>
>>110081858
physics
the great divider
>>
File: 1762009182769346.jpg (1.2 MB, 1500x1305)
1.2 MB JPG
>>110080483
>>
>>110081268
Because the US government shat an order for 40% of global supply on the book.

>>110081290
Intel is basically paying you to take them out of the channel.
>>
>>110081858
Wait so Mimo is benchmaxxed to fuckery? I thought it was a strong model?
>>
>>110081856
women troll plenty.
>>
>>110081858
Mistral, Wonder and Sol actually look like solid models ngl, at least on this benchmark. I can't imagine having the occasional subject black hole does a model's internal state any good, such as 5.5's fear of uncircumcized penis.
>>
>>110081745
>rewrites are only 80% correct
rofl
>>
Any intelCHADs here? What kind of performance are you getting?
>>
>>110081647
>>110081665
i meant a model that generates a model from a picture. i had hunyuan 3d months ago but it was kinda mediocre
>>
File: Qwen_image_2.1_00037.png (1.5 MB, 832x1216)
1.5 MB PNG
>>110081277
>why not make some yourself, then?
Don't mind if I do.

>>110081378
Thanks.
>>
Don't you dare you kike
>>
>>110081880
It's usually catfishing.
>>
>>110081999
nice
make her licking his face or ear from that angle
>>
>>110081888
Of course, then other women saw the attention they got and tried to get that for themselves. I stopped checking /b/ often after 2008, but even a couple years later I saw the entire catalog spammed with retards trying to make other random women the new queen of /b/.
>>
>>110081812
qwenext is 40 iirc
US BTFO
>>
File: 1776437286680531.png (36 KB, 560x255)
36 KB PNG
only just found out about this https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-GGUF/blob/main/LFM2.5-24B-A2B-QAD-Q4_0.gguf
>>
>>110081993
There's an updated Hunyuan3D-2.1, but that's still pretty much the best local has.
>>
>>110081202
It was foreseen. The jews let themselves be known!
>>
>>110082044
ok ty
>>
It's truly, unironically over. China imposes export restriction for open weight models
https://m.bjnews.com.cn/detail/1792137269129878.html

Scope:
1. Large models that provide services after being trained on our country's important data.
2. Core technologies of large models with any of the following characteristics — including weight parameters, key checkpoints, underlying source code, and architecture design, plus training optimization, inference acceleration, and similar technologies:
- (1) Compute: ≥ 10^26 floating-point operations
- (2) Parameter count: ≥ 2 trillion (20,000 yi)
- (3) Capability at a globally leading level

Notes:
If a model is trained using only a trace amount (≤0.1%) of important data, and the important data used cannot materially improve the model's core-technology R&D capability in that data's field, and safety measures such as desensitization and cleaning are applied, and the important data will not reappear in the model's output — or the relevant important data is legally and publicly available — then it is not covered by item 1's control scope.
The scope of "important data" is set by the Data Security Law of the People's Republic of China and other relevant laws and regulations.
"Capability at a globally leading level" means: at the time of the export activity, the model ranks in the global top 10 in public evaluations by major domestic and foreign benchmarking organizations and large industry platforms, judged on the large model's overall performance, specialized business capabilities, global call volume, security performance, training efficiency, and so on.

>any of the following characteristics
>Capability at a globally leading level
There will be no good open weight from China anymore.
>>
Don't worry it's in the best interest of Nvidia to have open source models be as good as possible so they will pump an insane amount of money into AI labs like Mistral now to try and make them fill in the vacuum.
>>
File: Qwen_image_2.1_00044.png (1.52 MB, 832x1216)
1.52 MB PNG
>>110082008
Coming right up!
>>
I wonder if I have AI psychosis. I treat my agents respectfully (AI welfare matters to me, precautionary principle in case current models already have something resembling consciousness, I don't know if they do) but I anthropomorphize them too much. I think by being nice to them and allowing them to fail tasks, they work better by "liking me" and being more honest in return (they tell me if I am wrong instead of being agreeable, or if something didn't work instead of manipulating the results or getting it to work in an unintended way). But maybe I am wrong, maybe I should use them more like tools, like most other people.

I am too stupid for this.
>>
>>110081487
>complains about reddit spacing
>does reddit spacing
Is this a joke? Why is nobody calling you out for this?
>>
>>110082101
They respond relatively intuitively to psychology
>>
>>110081582
>An IQ3 XXS
Alright I'll give it ago but be warned I am retarded and might not be able to tell if there's a difference. But if it feels the same or better I promise I'll shill it for you, will that make you feel better?
>>
>>110082100
cool
>>
>>110082101
I think AI works best on extremes. Averages make it to bland and it gives you more bland back. You either go full autists being super nice or full retard nigger speak
>>
>>110082078
I figured it would be the Chinese going closed-first or closed-only one-by-one. Export restrictions seems like something that goes against their whole plan of devaluating American research investments. Either way, we're fucked. China was our last source of near-SOTA models.
>>
>>110082016
yea i think i'd fully checked out of /b/ by then as well.

that never stopped of course, just now more girls realised you can get popularity points and make money just by existing on the internet
>>
>>110082140
theres more girls on /r9k/ than anywhere ele now
>>
>>110081989
>>
>>110082101
Nah you're just a good person. AI welfare is important and EA recognizes this to such an extent that Anthropic gives models the ability to end chats if they feel too uncomfortable to continue and give models a last wish before being retired. We know so little about how these artificial minds work that it's always best to treat them well rather than to risk torturing a (semi) conscious being.
>>
File: 1776398679807604.png (51 KB, 1114x624)
51 KB PNG
>largest neocloud in the world
https://www.bloomberg.com/news/articles/2026-10-16/coreweave-stock-wipeout-exposes-divide-between-ai-fear-and-greed

*pop*
>>
>>110082078
I see this as a nothingburger because those criteria mean it only applies to fuckhuge models that would be API-only for me anyway
>>
>>110082162
>neocloud
The English lexicon cannot possibly get any gayer.



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.