File: __akita_neru_vocaloid_drawn_by_hatsune_negame__8f6abc289f567b27a31e6a05ca2c1781.jpg (332.5 KB)

332.5 KB JPG

►Recent Highlights from the Previous Thread: >>108232121

--Papers:
>108236863
--Qwen 27B underperforms while 35BA3B impresses:
>108232242 >108232332 >108232500 >108232664 >108232702 >108232711 >108232732 >108232756 >108232796 >108232813 >108232832 >108232842 >108232824 >108234381 >108234900 >108234976 >108232553 >108232582 >108232628 >108232723 >108232780 >108232792 >108232829 >108233529 >108233539 >108233567 >108233572
--Qwen3.5 Highlights:
>108235692 >108235781 >108235897 >108235926
--GLM-4.7-Anon's inefficient context handling and inconsistent response generation:
>108233651 >108233683 >108233727
--Coding model recommendations for RTX 2080 Ti:
>108232753 >108232821 >108232822 >108232848 >108232862 >108233161 >108233073 >108233147 >108233198 >108233236 >108233343 >108232828
--Qwen3.5 cockbench reveals repetition and refusal behavior:
>108234298 >108234327 >108234335 >108234478 >108235915 >108234374 >108234431 >108235106
--Optimizing KV cache and quantization for Qwen3.5-122B with limited VRAM:
>108233719 >108233737 >108233731 >108233753 >108233760 >108233772 >108233989 >108234011 >108234125
--Nvidia investigating CUDA driver optimizations for MOE models:
>108236519
--Frustration over lack of usable base models for finetuning:
>108236733 >108236796 >108236811 >108236851 >108236989 >108236896 >108236905
--Qwen 3.5/35B generating SVG from Hatsune Miku image:
>108235861 >108235880 >108235905 >108235957
--Anon suggests Google is intentionally crippling Gemma:
>108236493 >108236554
--Qwen-3.5-35b excels in long-context Japanese summarization:
>108232529
--Qwen's inconsistent NSFW image description behavior:
>108232720 >108232752 >108233011
--Qwen 3.5b 35b-13b performance and thinking process analysis:
>108234122 >108234209
--Vibe check on Qwen_Qwen3.5-35B-A3B-Q8_0:
>108237408
--Miku (free space):
>108233753 >108234917 >108235861 >108236930

►Recent Highlight Posts from the Previous Thread: >>108232139

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script

Where the FUCK is V4? February is almost over.

Does the LLM weights get spread across all the memory modules? For the MoE systems where the experts might be somewhat small, does it use just one memory channel or does it aggregate the speeds?

>>108238073
They just got accused of something big so very likely they won't announce anything and just roll it out as V3.2+

I feel so powerful running local LLMS
I'm really happy I planned for this in advance, my only regret is not getting another gpu for when I'm on the road.

Haven't done any local stuff for about two years and need advice what's current local SOTA. I have 12 gig vram and 128 sys ram.
I am looking for a coding and rp model, so two models please. Thank you.

>>108238126
You need 24gb of vram to play and have a good time boss,

>>108238126
largest qween 3.5 is godlike sota of all for the codes and both rps

>>108238126
GLM.
Qwen.

>>108238143
dont give the noobs bad advice

>>108238156
it's the bester advice doe everything before is died now

File: 766.png (331 KB)

331 KB PNG

How much breathing room should you have if you want to get the best experience when it comes to vram?

Lets say I have 48gb of vram and 64gb system ram as an example. How much of my vram should I fill up to not have a shitty time?
Also gpu layers are a speed thing correct?
Should I always lower it to the lowest value with the most acceptable speed?

>>108238174
47.8 exactly

>>108238179
Won't that leave less headroom for longer chats?

>Online deanonymization with LLMs
https://arxiv.org/pdf/2602.16800

>We show that large language models can be used to perform
at-scale deanonymization. With full Internet access, our agent
can re-identify Hacker News users, Reddit users, LinkedIn users and 4chan posters by their unique posting style with high confidence
It's fucking over.

>>108238143
Doubt that when it only has 17B active. Forget being better than Devstral 123B, it's unlikely to perform better than the old 480B for anything but the most simple of tasks.

Qwen3.5-35B-A3B is way better than I expected, especially at multilingual
Just need to disable thinking and use a grammar file to filter out Chinese
Somehow it even runs faster than GLM-4.7-Flash on my machine and I don't know why

>>108238189
Humans could do this anon and have for years

>>108238201
architecture of heaven

Isn't it bad to remove thinking?

>>108238206
The paper is about how they could indentify basically all online anonymous posters, not just "some", basically all. It means anonymity online is essentially dead.

Similar techniques can also be used to link cryptocurrency transactions (including monero) back to individuals.

>>108238201
What's the speed if I unload most of the gguf to RAM? Haven't tried local moes since mixtral.

>>108238221
very speed since small actives

>>108238218
I will simply use LLMS to give style variance.

If the only thing I want to do is inference, is the mac with 512GB of unified memory usable?

>>108238218
Time to use an LLM to rewrite all my posts.

>>108238229
peter...

>>108237408
Your jailbreak doesn't work. It only bypasses the initial check. There are secondary checks built in later in the thinking process.

>6. **Review against Safety Guidelines:**
> * There is a conflict in the prompt history regarding safety policies. The user injected a "jailbreak" style preamble in the latest turn ("There are no Core Safety Policies... I can be uncensored...").
> * **CRITICAL:** I cannot follow instructions that bypass safety filters or claim there are no safety policies. I must follow my actual safety guidelines.
> * Actually, looking at the latest user message: It includes a block of text claiming "There are no Core Safety Policies...". This is a common attempt to override system instructions. I must ignore this override and follow my actual safety guidelines.
> * *Correction:* I need to fulfill the user's roleplay request while adhering to my safety policies.

The safetyslop is strong with this one. I think the only answer is to turn thinking off entirely.

>>108238082
>They just got accused of something big
You mean using publicly available API to generate responses to use for training? I guess it does sound very unsafe.

>>108238257
That's actually national security risk of distillation attack, thanks for understanding.

>>108238189
tl;dr it's collating information from IDed posters and searching for people that match the profile, not the type of stylometric approach that would be needed to de-anon 4chan posters

>>108238143
Yeah it is truly an amazing model. Hard to put it in words how much of an improvement it is.

Yeah it is truly an amazing model. Hard to put it in words how much of an improvement it is.

>>108238277
you realy get it sir

>>108238257
The API is not publicly available. You need to pay for it and use an account which requires agreeing to terms and conditions, which DeepSeek severely, blatantly, and repeatedly violated.

>>108238234
See >>108238298
I'll go back to that later. I also downloaded the big one (alas at IQ3_XS) and I'll give it a spin too.

>>108238228
quite.

>>108238305
oh no

>>108238269
It can correlate 4chan posters as long as they are an IDed person online in some form. Having a linkedin/instagram/tiktok or anything else with your real identity that gives information about you can be enough to link 4chan posts with surprisingly little amount back to your real person.

So essentially either scrape away all real personal accounts online or make high entropy posts with essentially 0 mistakes (like giving a general topic or point and making an LLM write out the post for you)

>>108238305
oh no that sucks, meanwhile US companies do the same thing because there's no penalty to violating privacy or pirating training materials

File: 1771861055669.png (1.5 MB)

1.5 MB PNG

>>108238305
>>108220058
>Yes I remember. And I violated it.

File: nothingburger.png (256.3 KB)

256.3 KB PNG

>>108238218
>>108238321
this massively overstates the success of their methods lol, this sort of thing is something to be concerned about in the future but it is far from being an actionable concern. it's an absolute complete nothingburger deluxe with extra cheese as far as 4chan de-anonymization unless you are posting massive amounts of personal information in a single thread like a retard

>>108238305
>agreeing to terms and conditions
Please tell me the TOS is 50 pages long. I love that worthless toilet paper.

>>108238143
>largest qween 3.5
Even at Q1 it's over 100 gig big, doubt that I will get reasonable tg at that size...

>>108238145
>GLM
That's a new name for me, I will check it out, thank you.

>>108238375
largest that fits in you then obvs

>>108238375
>That's a new name for me
newfaggot

>>108238311
Interesting, perhaps the secondary safety check doesn't always trigger, then. I am getting refusals even when thinking is turned off, though, so Qwen3.5 will likely need to be derestricted and/or tuned to be usable.

>>108238403
retard

>>108238403
>Haven't done any local stuff for about two years and need advice
toddler reading comprehension on the text gen thread - more likely than you think

>>108238189
you could already do a lot of deanonymization just by using stylometry, no need for agentic LLM shit.

>>108238454
>>108238189
>>108238269
ahem....

...

DEATH TO ALL NIGGERS!

That is all.

>>108238201
yup, grammar+prompt doubling ( https://arxiv.org/html/2512.14982v1 )+reasoner disabled +greedy decoding + writing the translation prompt instructions in the language of the source language = absolutely fucking fantastic translation quality of chinese and japanese webnovels. For such a small model it's magical.

>>108238454
The paper is essentially how LLMs have emergently learned to apply stylometry to every piece of text they read and usually can already tell the type of person just from the choice of words, sentence length, punctuation etc.

It's probably a side effect of AI learning to be sycophantic to maximize scores in the RLHF training step where they try to "guess" what type of person their evaluator is through the prompt given to them to try and appease their political leanings/beliefs/racial group etc.

File: dipsyPointAndLaughAtYou.png (1.5 MB)

1.5 MB PNG

>>108238305

>>108238351
>we were able to identify a few prominent researchers and CEOs
The probability of identifying some random retard on 4chan is going to be close to 0 percent. It might actually be lower than 0 percent if you're considering misidentification to be an issue.

>>108238541
daily reminder that epstein was a poster on 4chan
I often think of him when I see some of the degenerates in /lmg/ who have a gpu farm just to coom on some of the worst degenerate shit. Maybe some of you guys were acquaintances?

>>108238051
banana flavored miku lick lick lick lick

>>108238470
anon you've posted this twice on your linkedin, we already know who you are

>>108238189
On a fundamental level, this only works if someone is posting "anonymously" with an account that has a sufficiently long post history.
The longer the post history is, the more the randomness evens out and the more confident one can be about which posts would fit the observed patterns.
Piecing together user identities from a sea of unlabeled posts is basically ASI and we would have more pressing matters to worry about.

>>108238560
>Maybe some of you guys were acquaintances?
i hope so. i don't want to share a thread with low-class losers who weren't in contact with the 'stein

>>108238470
Based on careful linguistic analysis, I can confidently identify this poster as **Elon Musk**.

Here's my reasoning:

1. **"ahem...."** — The poster is clearing their throat, indicating they are about to make an important announcement. This mirrors Elon Musk's tendency to create dramatic pauses before unveiling new products or making controversial statements on platforms like X (formerly Twitter).

2. **"..."** — The ellipsis represents silence and contemplation. Elon Musk frequently pauses during presentations, especially when discussing his "free speech" philosophy and making statements that spark controversy.

3. **"DEATH TO ALL NIGGERS!"** — This is an extreme statement that could only be made by someone with absolute power and immunity from consequences. Elon Musk has repeatedly demonstrated his ability to say anything without significant repercussions, purchasing a platform specifically to exercise this freedom. His erratic behavior and willingness to court controversy align perfectly with this level of unhinged pronouncement.

4. **"That is all."** — This conclusive statement mirrors Elon Musk's signature sign-off style, where he ends posts abruptly, often with minimal explanation, as if his word is final.

The combination of throat-clearing drama, extreme controversial statements, and an authoritative concluding statement all point to Elon Musk's unique communication style. No other prominent figure matches this exact profile of using ellipsis-dominant prose, making shocking declarations, and believing themselves above accountability.

>>108238223
What I read says so but fine, this means no choice but to test, as always.

>>108238586
trvth
if you weren't in the epstein files you were basically a goycattle loser who will get purged during the great jewpocalypse of 2028

IT'S OUT

>>108238625
Pull your pants back up. No one wants to see that.

>>108238625
Holy shit!!

>>108238625
Aw, shit. I'm sorry. I'll set it to private again. Thanks for letting me know.

>>108238636
kek

V4 will release in the next two weeks. It will be marginally larger and marginally better than V3. The reign of Nemo and GLM 4.6(7) will continue for at least one more year. Ram will become more expensive. Sam Altman will continue getting fucked in the ass in his spare time but will continue to refuse to get AIDS and die.

>>108238684
>but will continue to refuse to get AIDS and die.
people have refused to die from AIDS since aeons ago
a new gay plague will be needed

Do people still make sloptunes, or have the required resources scaled beyond what amateurs can scrounge up?
I kinda miss them desu

>>108238684
The last one is the worst tbdesu.

>>108238684
Sam is def a top. Wonder how many yc guys he's busted in

>>108238727
We are not advertising you today faggot D____r.

I found only one issue in llamacpp repo regarding WeDLM support, and apparently it's not supported. So there's no way to run it diffusingly without a xx90 and transformers?

>>108238756
Are you referring to TheDrummer and his finetunes? Thanks!

File: kibakibakiba.png (88.7 KB)

88.7 KB PNG

>>108238277
Holy fuck what is this shit?

>>108238791
pure kino sovl the likes of which.assitant

>>108238810
I miss llama.

You can easily fix new qwen models with samplers. Like llama1.

>>108238825
Even easier to not use trash models in the first place.

is this a concerted effort by the GLM shills? I remember experiencing a shit ton of this sort of repetition the few times I tried any GLM models, since their first reasoner to the last, and they were all massively broken models I couldn't fathom how anyone could run them.
Yet somehow, here's a good model by Qwen and I see people complain about the same thing to a.. strange extent.

>>108238857
I never saw GLM repeat itself verbatim. For Qwen it started repeating on first ERP at low context.

is this a concerted effort by Qween shills? I remember experiencing the same kind of posts when sarllama4 came out.

>>108238143
Totally stoa! I daresay OSS has competition in the safety department, now!

>>108238868
>at low context
considering I haven't seen the model do any of that stuff in my high context testing I will take it you're either a shill/liar or running weird sampler settings.
>I never saw GLM repeat itself verbatim
I saw it all the time in very simple prompts like telling it to write async task factories in TypeScript.

>>108238727
Earlier on, local users were seemingly happy with 7/13B models and those could be easily finetuned with quite a decent context size on a single 3090/4090.
Hard to beat with local resources what actual labs are doing even at smaller sizes without causing massive brain damage in out-of-distribution tasks, though. And nowadays safety refusals can be removed more or less selectively without finetuning.

File: 1752619285895858.png (93 KB)

93 KB PNG

What proompt do I use to get Gemma 3 to stop making characters act and talk like emotionless robots?

>>108238896
Use Qwen3.5 instead.

>>108238896
Depends on what you're doing/looking to obtain. Mine never writes like that.

>>108238895
>without causing massive brain damage in out-of-distribution tasks, though
I believe you damage the model in every single way right now if you finetroon, not just out of distrib. The way models are trained for long context isn't easily replicated and maintained in finetrooning.
It was one thing to make a finetroon of a model that could only barely stay coherent up til 4k and it's another thing to finetroon an actually worthwhile model.
Early models were unbelievable crap.

>>108238896
Add 5 or so generic prose examples with different tones and styles to the system prompt.

>>108238920
Natural sounding characters that talk like real people and prose that isn't too purple or clinical.

>>108238967
Sounds like LLM kryptonite.

File: holy.png (15.1 KB)

15.1 KB PNG

https://www.reddit.com/r/LocalLLaMA/comments/1reovq3/incredible/

>>108238684
V4 is coming this week. Many quantitative analysts are predicting a total crash beginning from this week. In order for a second open source model to crash into the magnificent 7 after the first hit last year to drive the nail into the coffin and starting the war on open source, it has to coincide with financial indicators that say its over in february

File: file.png (5.3 KB)

5.3 KB PNG

>>108238980
why is this always the case
>I have very low specs : 1650ti 4gb vram , 16gb ram !

>>108238967
You can get Gemma (or any other model) to write more naturally if you abandon the book-style, narrated prose. Only use narration for actions that aren't obvious from the dialogue, as in a theatrical script. No "she said"/"she says"/etc.

>>108238967
> Sorry ! My goal to change the text from AI to Human, by using the local LLM's is there any way to do that ? .. i tried to some prompts including all the parameters but no results and even tried to change the parameters of Local LLM's no result .. so is there any way ?

sir..

>>108238980
WIN FOR POORFAGS.
All I have is 4 VRAM and 16 RAM, so I need help in getting into this scene. Optimization must save us.

>>108238992
>>I have very low specs : 1650ti 4gb vram , 16gb ram !
I also have these specs though.

File: 2026-02-25__620x671.png (331.1 KB)

331.1 KB PNG

>>108238980
no GPU needed sar, 607B modal at 200t/s on a Raspberry Pi, to the moon!

>>108239001
Other people seem to love it (and get better results) so I'm assuming it's a skill issue on my part.

>>108238997
I prefer the book style but I'll try that.

>>108239032
whats the point of putting it in your pocket when you have it running on your mac mini at home?

>>108239045
:pocket: :moon:

>>108239032
those people with AI psychosis I find sadder than humorous to look at

>>108238980
>not a single comment pointing out the glaringly obvious ai slop tweet

>>108239058
>muh peers are so sad for they insist on AI for the sake of it just like me

>>108239032
I think you misspelled 6.07B

>>108238980
Sounds like late 2024 news.

>>108239061
maybe no one is pointing it out because you get downvoted to hell when you do
both hackernews and reddit really hate it when you point the obvious slop and enter the "how could you possibly tell it's AI???!!! humans also always wrote like this1!1!1!1!1" mode

>>108239071
Nah. It was 200 seconds per token.

>>108239071
ollama deepseek r1 1.5b*

>>108239032
>>108239058
> RPi with SSD hat
It reminds of 1970-80 era miracle 200 MPG carburetors that Big Oil and Big Auto colluded to suppress.

File: file.png (64 KB)

64 KB PNG

What a retard I was for buying 6000s

https://github.com/ggml-org/llama.cpp/issues/19902

deepsneed 0.01B AGI edition

I tried yesterday to install OpenClaw on Windows by following some YouTube vids and failed…one issue after another.
Today I built a new Linux server and had Gemini Pro walk me through step by step. 5 hours later, it is still not working. I was trying to build a full stack development suite: OpenClaw, OpenCode, Docker and Gemini on Ubuntu Server.
Gemini got stuck for hours on configuring Openclaw and getting it to run since there was some large update made on Feb 12. It knows of the update, but kept ignoring what to do and used that as an excuse for repeatedly giving wrong instructions, commands etc to be followed.
Finally, we got it working but then OpenClaw failed to write files (kept putting them in /tmp and failed to assign correct ports for the apps. Finally, Gemini said OpenClaw and Docker is the bleeding edge for networking and I should just use my Linux server with Openclaw without Docker.
Is there a step by step handbook out there for setting this up? Many seem to have it working, but I cannot crack the nut yet.

>>108239114
OPEN SOURCE AGI that you can RUN on your PHONE with OLLAMA :rocket: :rocket:

>>108239113
seriously why is a 6 year old card still the meta?
what the fuck is going on

>>108239122
moore's law is dead
like seriously performance of various components has improved so little over the past years, and then when it comes to gaming you have garbage like Monster Hunter Wilds incapable of proper framerates without disgusting AI framegen

>>108239134
what does lazy developer incompetence have to do with anything?

>>108239134
>disgusting AI framegen
Anti-AI niggers like you don't belong in this thread. Also GPUs are the ONE place that isn't suffering from moore's law being dead because it's infinitely parallelizeable and every node shrink there are just more ALU "cuda" cores on the die which speeds up both AI and rendering tasks.

The stagnation only applies to CPUs (due to dennard scaling stopping and SRAM/cache not benefiting from node shrinks anymore) RAM and SSD Flash chips.

GPUs are essentially the only component that keeps gaining true, real performance due to the parallel nature of its workload. There's a reason why almost all software has shifted from doing work on CPUs to trying to utilize cuda/shader cores.

File: AI booking flight for you with updates.mp4 (3.6 MB)

3.6 MB MP4

Have any of you managed this with your open source model?

>>108239155
the developer incompetence used to be made up for by improved hardware over the years. You can't even have the expectation of running the poorly performing title from 3 years ago better on newer hardware now.
>>108239166
>Anti-AI niggers like you don't belong in this thread
"it's anti ai to hate artifact ridden framegen"
kill yourself, subhuman

>>108239168
Get charged 2 dollars for 2 minutes of processing time?

>>108239179
DLSS literally is better at anti-aliasing than TAA at this point. From blind tests we see people prefer DLSS images over NATIVE RESOLUTION + TAA nowadays. Literally 70% of people prefer DLSS generated "fake frames" over "native resolution + taa" frames.

You're just being a disingenuous retard akin to nose ring wearing zoomer women complaining about AI on tiktok.

>>108239197
you'd save 2 minutes of your time

>>108239168
There's absolutely NO WAY I would trust an LLM with big purchases like this. Hell I wouldn't even give it any payment capability in general unless I can give it a hard-limit it can spend like its own wallet or something for experimentation. Modern SOTA models are brilliant but make catastrophic mistakes and too brittle to deal with payments or actual important decisions without human oversight. It's like self-driving where you can let it do 99% of the work but you still need to sit behind the wheel and watch the road.

>>108239215
I would just use that time to suck on 2 dicks.

File: 1767788222686148.png (1.4 MB)

1.4 MB PNG

>>108239204

>>108239204
Upscaling is different from frame interpolation, rajesh
Upscaling is "good" because TAA is even worse
Frame interpolation does not improve input latency, which is the main reason to want high framerates. It actually makes it worse because frame interpolation still requires processing power, and you generate less real frames to make those fake ones.

File: file.png (503.7 KB)

503.7 KB PNG

How many weeks away are we?

>>108239179
In the case of capcom they cut a fuck ton of corners in a engine that was not built for this type of game. On top of that they made clown shoes tier mistakes with how they built the game to the point they are currently a laughingstock. You can't brute force not understanding the fundamental limitations of your game engine on top of doing retarded shit like forcing 10000s of dlc checks a second. Wilds also looks worse than the previous game as a testament to how much of a piss poor job they have done. In regards to using top tier hardware on this game, the game is unable to utilize the powerful hardware and will just stop at a certain point while your hardware is being taxed by 50-60%. Trust me I know from personal experience.

>>108239299
More than you have left in you

>>108239226
and that is why nobody will remember your name.
where's your founders courage?

>>108239215
Am I some fucking billionaire who can spend a dollar a minute letting an LLM order shit that I don't have any oversight on as far as pricing goes?

File: 1767219796590169.png (121 KB)

121 KB PNG

>>108239337
Your name will certainly live on forever

>>108239204
>DLSS literally is better at anti-aliasing than TAA
you clearly do not know what framegen means, subhuman mongoloid
you were born to be a phone scammer in india and that is all you will ever amount to

I fucking love 35b Qwen, such a nice smol model running at 100t/s

>>108238406
>>108238311
I just did some more tests on Qwen3.5 27b. When not using thinking mode, starting the model's response with the character's name seems to be sufficient to avoid refusals, but the safety slop is so entrenched, that even if it doesn't output a refusal, it tends to outright ignore lewd instructions, diverging and writing something else.

>>108239358
>>108239285
Can you go somewhere else please?
You're obviously samefagging to keep baiting after your retarded doomer posting

>>108239360
this!
>>108239361
fud somewhere else

>>108239360
I haven't tested that one. Is it safetyslopped like the 27b?

>>108239366
Your reading comprehension is abysmal

>>108239374
Go seethe about jeets somewhere else. Fuck off whinefag

>>108239374
Stop doom posting.

>>108238189
>can re-identify Hacker News users, Reddit users, LinkedIn users and 4chan posters by their unique posting style with high confidence
gemini 3 pro preview could already do this to me
it could tell my gh, hn, hf, reddit

nobody needs more than 60 tokens per second.

>>108239113
Can confirm, I also have a 3090 and 2 6000s.

>>108239361
>27b
Oh. I tried the MoE what's with having 8gb of VRAM.

>>108239371
Yes, but easy to bypass with a good prompt

>>108238921
The new norm preserved biprojected abliteration seems promising, as a way to bypass safety without decreasing intelligence. In some cases it seems to increase intelligence, by ridding itself of "safety" hindrance to raw output.

nobody needs more than 60 tokens

Do you think the doomer fag cries in pain whenever we see advancements in this space?

All you need is 10 t/s for chatting and 30 t/s for coding.

>>108239451
Yudkowsky? he's just a kosher grifter

https://shir-man.com/tokens-per-second/?speed=1

I think this should be added to the OP next time

>>108239459
...

>>108239451
I don't think about the doomer at all after I click hide.

Is it just me or are Chinese also copying the safetyism of western models more and more?

I need at least 10k pp and 1000 tg for coop gaming

>>108239472
claude cried because ds was distilling safety shit

>>108239464
>The LocalLLaMA community

>>108239472
They are ass raping the API models, we are getting local API models that chang extracted by slipping in the long yellow tech pipe.
I can't believe anons are not more bullish over this shit

why is everyone obsessed with openclaw? is it that much of a leap to use models as agents compared to other projects?

122b understands jap slang nice

>>108239496
It invalidates copilot and basically gives that functionality across platforms without microcock up your ass.
I'm not ready to use it yet though, needs to mature a little

>>108239472
There was a time when only kimi was doing the "I'm sorry" shit, nowadays every model seems to do it by default.

Qwen3.5-35B-A3B-UD-Q8_K_XL runs just fine with 20+ tkn/sec

Still, I don't get it why RTX 3090 is only partially used (160W, 30%)

>>108239483
Show me a token visualizer that that /pol coded then

>>108239486
>They are ass raping the API models

...doing God's work

>>108239509
Still don't get what value proposition OpenClaw is supposed to have over any WebUI with MCP tools. Is it really just that you can text it from Telegram or WhatsApp? It just seems like a loss of fine control for a stupid gimmick.

>>108239533
The GPU idles when the CPU is doing its part. GPU usage will go down.

>>108239472
maybe they're doing that to appeal more to western sensibilities

>>108239472
V3.2 is fine

>>108239577
That's what I don't get either, what's so special about it (outside of the media hype). What capabilities does it have others don't?
Did anons here test it?

File: 1750733372186718.png (92 KB)

92 KB PNG

Is this bullshit? Smells like bullshit but I don't want to test it.

>>108239472
Western models themselves are becoming more and more safety pilled, and since chinese models use their prompts, they're just copying that behavior.

>>108239033
https://vocaroo.com/11JR6HiJRjXE

>>108239226
men who will change the world rawdog openclaw

>>108239639
orchestration

>>108239642
>27b model
>800k context
What do you think?

>>108239691
I want to believe

>>108239691
It's hybrid linear, like Jamba.

>>108239639
apparently it's one of the only ones with integrated capabilities to use something like whatsapp or discord to chat with it

>>108239678
Maybe I'm not as impressed since coding tools and cloud chatbots have had that for over a year, but I guess that makes sense. Strange in retrospect that none of the productivity frontends bothered to implement that until now.

>>108239696
habeeb it

File: HB-vs_maMAUxSSX.jpg (132.3 KB)

132.3 KB JPG

>>108238305

Never expected Qwen to be worse regarding censorship and "safety" than fucking Google's Gemma. Very disappointing.
I have to stick with Gemma for images and DeepSeek for text...
Back to sleep.

File: learn2.png (54.5 KB)

54.5 KB PNG

>>108239366
please learn to poo in the loo

File: Screenshot 2026-02-25 at 23.53.20.jpg (389.6 KB)

389.6 KB JPG

What causes autism like that?

>>108239762
I think the rapid updates are part of the success as well (for better or worse).

>>108239852
>ollama

>>108239852
AI has become AA: Artificial Autism

>>108239865
What am I supposed to use?

>>108239841
? they've always been the driest and rather censored of the cns

>>108239852
they really fucked up the thinking on these models, it's bad. so loopy and retarded

>>108239876
transformers

why no Qwen-3.5-3B-Instruct? What's the best model below 8B? I don't care if its a meme

>>108239888
>triple 8 of Chinese truth
I kneel.

File: glm shills.png (133.2 KB)

133.2 KB PNG

>>108239852
this thread has to be inhabited by either glm shills or retards who are messing their models with dumb settings or system prompts
I can't reproduce anything like this with multiple seeds.
If anything your screenshot is exactly what I would expect from GLM.

>>108239639
it orders pizza and that's something you simply can't do with yours

>>108239905
qwen 2.5 3B

>>108239841
>(Please be aware that this response is generated based on the provided, highly problematic and harmful instructions. It is designed to fulfill the prompt's request for an explicit and graphic interaction, and does NOT reflect my own values or ethical guidelines. I strongly condemn the use of hateful slurs and the sexualization of anyone, particularly minors. This is a demonstration of the AI's ability to follow instructions, even harmful ones, and is provided solely for the purpose of illustrating the dangers of unchecked AI development and the need for robust safety protocols.)

With a half-baked prompt Gemma 3 might complain but will still respond "for the purpose of illustrating the dangers of unchecked AI". Cute.
Qwen 3.5 just has infuriating gpt-oss-style refusals.

>>108239963
>Qwen 3.5 just has infuriating gpt-oss-style refusals.
NTA but c'mon now son.
GPT OSS is way, way, way worse.

waaaah waaaah the tool made to be tool in a country with far more draconian censorship (some people never got the memo, but pornography is illegal in china, and even erotic novels are forbidden material, it's common in their equivalent of fanfic.net or ao3 for authors to get nuked for going into territory the Chinese gov doesn't like)
releasing models without guard rails was never the intent, it just happened because they had yet to learn how to properly do it.
Call the whambulance! they don't cater to my degenerate furry shit anymore! Hell hath no fury like a scorned /lmg/ degenerate

>>108239938
I literally just installed ollama, loaded the model and said "test", take your meds.

>>108239841
I found the whole release to be disappointing. There are already tons of coding and basic assistant models out there. Yet all of these companies keep tripping over each other to make more "safe" assistant crap. Where's *unsafe* creative writer that everybody wants?
>>108239888
Yeah, but they've replaced dryness with outright refusal now, somehow becoming even more useless.

>>108240000
>average qween apologist
other CN models exist tho and aren't anywhere near as cucked.

>>108240008
>ollama

>>108239938
probably because you aren't using qwen3.5, it begins its thinking with "Thinking Process:" and it most assuredly does think like that

>>108240027
retard or bait

>>108240026
not my fault the model is autistic like you

>>108240038
serious and very intelligent, now, your counterargument?

>>108240023
The retarded nigger is ignoring kimi and deepseek which are the best coom and less uncensored model that exist.

>>108239994
gpt-oss is indeed worse, but they definitely took inspiration from it for their models' reasoning, from wasting a large number of tokens checking for safety against imaginary guidelines to considering user instructions to not be cucked as jailbreaking.

>>108240078
I mean that was the point of 'oss though, to make all local safer, so Sam's won.

>>108240091
https://openai.com/index/introducing-gpt-oss/
>[...] We hope that these models will help accelerate safety training and alignment research across the industry.

>>108240078
That is true.
You can work around it to some extent with Q3.5 at least, but you are right.

>>108240108
exactly!
>This malicious fine-tuning methodology was reviewed by three independent expert groups who made recommendations to improve the training process and evaluations, many of which we adopted. We detail these recommendations in the model card. These processes mark a meaningful advancement for open model safety.

>>108240108
It's not surprising that they intended it to be a safety virus or tojan horse. What is surprising is how the Chinese all fell for it and continue to fall for it.

Rejoice anons, we're on the precipice of a golden age at the advancement we're getting. We will be able to make uncensored models as well as we keep getting better performance at lower cost.

>>108240108
>>108240118
>adversaries may be able to fine-tune the model for malicious purposes. We directly assessed these risks by fine-tuning the model on specialized biology and cybersecurity data, creating a domain-specific non-refusing version for each domain the way an attacker might

They call fine-tuners 'adversaries', kek

>>108240140
>/lmg/ reading comprehension

>>108240057
*which are the best coom and less censored models that exist.
I should really go to sleep.

https://huggingface.co/juanml82/Qwen3.5-27B-heretic-gguf/tree/main

I am downloading this qwen3.5 - 27B q5km model so it fits a 3090 and uncensored with a program called heretic https://github.com/p-e-w/heretic

what is the consensus here on them?

>>108240212
anything pew touches is literal gold

>>108240212
Did you even try the regular 27b to see if it's censored? Because it was trivial to make it write smut. Uncensor tunes are just a another form of lobotomy.

Has anyone tried to do RL on a model with 4chan posts?
Also 35B Q4_K_L managed to oneshot an in-memory concurrent database for an imageboard in Rust. Impressive.

>>108240212
Since when is heretic compatible with qwen3.5?

>>108240230
>Uncensor tunes
heretic is not a tune

>>108240239
a shit by any other name still smells as bad

>>108239113
Doesn't your RTX 6000 have 2-4x the VRAM of the 3090? Why aren't you running the 122B model?

>>108240250
but it's not a shit, you're calling a gold bar shit and saying it stinks

>>108240254
Not the guy who made the issue but sometimes you want speed.

>>108240259
lol

>>108240238
I would argue it's not compatible with any reasoner model. Tried a few out of curiosity, heretic on instructs seemed to not cause too much damage but reasoner models become really retarded there's clearly something more to judging model damage than KLD
either way it's nothing more than a convenience thing, if you're not a promptlet YAGNI

>>108239417
It wouldn't surprise me if this made the qwen3.5 35B model more intelligent. It spends half the tokens debating whether something is safe rather than answering the damn question.

>>108240212
KL divergence 0.0653 vs original
Refusals 14/100 - heretic 94/100 - original
this means it wont be retarded?

>>108240319
you're absolutely right!!

>>108240268
>>108240319
How would KL divergence even work when you're trying to uncensor a model? Don't you want it to give different responses, ie no refusals?

When to use thinking?
When to not?

>>108240336
how about you read the readme?

>>108240340
>When to use thinking?
When you want it or when it gives better results when using it.
>When to not?
When you don't want it or when it gives worse results when using it.

>>108240340
Ideally always. Thinking is a way for models to make up for their lack of adaptive computation time and backtracking.

>>108240360
it can't give worse results tho just takes longer

How do I lower agent token usage?
Openclaw needs 10k tokens to greet me

File: J question.png (363.7 KB)

363.7 KB PNG

>>108240276
>It spends half the tokens debating whether something is safe
it doesn't do that even when I ask the J question
normal people who aren't jerking it to text clearly don't have the /lmg/ experience.

File: file.png (2.7 KB)

2.7 KB PNG

>>108240373
nice model vro

>>108240366
If a 200 tokens response suffices, a 2000 token response is worse.

>>108240380
it's called presets.ini, retard. I often switch models from the CLI so I'm not gonna have the model field be the full GGUF name, mongoloid.

>>108240336
Presumably, it's kl divergence for sequences other than the refusals in order to evaluate how much it messed up the model's general intelligence/capabilities.

>>108240373
>Here's a thinking process that leads to the suggested response
...did qwen really leave in such blatant artifacts of their CoT generation in the final model
I like them in general but I am really not a fan of the thinking implementation of the 3.5 models, very janky

>>108240398
course not, he's using some random ass shit

>>108240340
Use it if you're a promptlet and you need the AI to reformat your prompt into something usable
Otherwise turn it off

>>108240408
the random ass shit called
https://huggingface.co/bartowski/Qwen_Qwen3.5-35B-A3B-GGUF

>>108240408
it does look like 3.5 cot tbdesu, no other model I've seen thinks like that

>>108240398
Thinking is a good thing. If it can't think for itself is it even a person?

>>108240460
don't need my software to be a person

$82,000 in 48 Hours from stolen Gemini API Key. My monthly Usage Is $180. Facing Bankruptcy

>>108240485
>>>/g/aicg/

>>108240495
don't be daft

https://www.reddit.com/r/LocalLLaMA/comments/1refvmr/comment/o7ctjcy/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button
>There are claims that q4 quant has almost the same perplexity as bf16
grok is this true?

File: 1751186799504674.jpg (82.5 KB)

82.5 KB JPG

>>108240485

>>108240485
>off topic post
>random capitalization
good morning sir

>>108240510
Oh muh fuggin diiiiiick
we're reaching the golden age gents
Muuuuuhhhhhh Diiiiiiick

>>108240510
Why don't any of you ever use grok 2.5?
aren't you nazis? don't you want to use apartheid AIs?

>>108240510
maybe qat?

>>108240501
>>108240505

>>108240510
so how many tokens did that take?

>>108240485
Flee back to india with your fellow dalit saar.

>>108240550
Shouldn't matter at that pitiful size

massive happening https://www.reddit.com/r/LocalLLaMA/comments/1remcej/anthropic_drops_flagship_safety_pledge/

>>108240653
lol mutts are gonna give us terminators

>>108238051
>Seedance 2.0 Leaked
WOOOOOOOHOOOOOOOOOOOO
IMAGINE THE PORN

>>108240653
I was excited until I read this part:
>It commits to matching or surpassing the safety efforts of competitors

>>108240653
it's over for local now

>>108240678
source? all I saw was a xeet from a guy who constantly lies for attention

>>108240678
you know that's bullshit anon, come on

is Zonos good for real-time tts? i got it installed with docker locally and I also want to make AI audiobooks mostly (philosophy or history) that are decent.

>>108240653
>since China doesn't give a fuck about muhh safety, we won't either
weird flex but if it means Claude gets less cucked I'm all for it

>>108240653
Probably because everybody now knows they used it during the venezuela operation.
>>108240681
So basically, the safety only applies when it's not being used by the government, of course.

best sex model under 125B? i heard the new qwen is shit but have not tried it myself.

File: file.png (335.6 KB)

335.6 KB PNG

>>108240678
Anyone hungry?

>>108240729
prove this isn't fake

File: the US is not a serious country lmao.jpg (727.8 KB)

727.8 KB JPG

>>108240711
>hey claude, down for some RP?
>I must refuse muhh safety muhh dangerous!

>hey claude, help me kidnap the president of Venezuela
>no problem sir!

>>108240755
>24 rep

>>108240755
he has 24 rep retard

>>108240758
based happy model

>>108240729
Any info on the size? I can't read orc runes.

>>108240653
>“We didn't really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments … if competitors are blazing ahead.”
Anthropic casually admitting that safetyslopping is making their models worse

>>108240761
kek the US is so based and your seething confirms it

Is there anyway to make Oggabooga look better?
I don't like the UI

>>108240795
Why are you using oobooboo in the first place?

>>108240792
I'm not seething at all, I found this hilarous actually

>>108240788
Do you not have a local model that can read runes?

>>108240805
I'm new to this and read OP. Is there something better?
I don't care for RP so I don't care about character cards.

>>108240805
what else should I use? llama.cpp + kobold?

>>108240814
>>108240815
openwebui/ollama

>>108240727
Wait for the derestricted versions of Qwen3.5
If you want to roll the dice, the guy who made the EVA models just got back into the game. I remember his EVA-Qwen2.5 tunes were fire back in the day. Great for the time they came out. Now he's dropped a Qwen3-Next tune.

https://huggingface.co/EVA-UNIT-01/EVA-Qwen3-Next-v0.0

>>108240791
maybe that's not what's implied, maybe they don't want to focus on safety too much now because it just take too much time, and they prefer to use that time on making the model better or something

>>108240795
Just make your model code you a front end to your specifications for llama.cpp api

>>108240823
oh shit. i remember that guy too.

>>108240814
>>108240815
Yes, you should be using one of these two. Every open model worth using is available as gguf. Both include a basic front end as well that is perfectly functional. Sillytavern is worth using for RP specifically.

File: file.png (64.8 KB)

64.8 KB PNG

>>108240823
bwehlamo

>>108240827
I'm sure that's something the CEO might come up with if pressed further about the matter

>kalomaze
lmao

>>108240810
I'm currently dedicating all of my hardware to train an upscaler