migu :3

>>108281695
sex

what's the best model ever?

>>108281704
the original pre-lobotomy c.ai model is still unmatched in terms of pure soul

>>108281704
davinci 003 for writing stories, this shit was absolutely insane

>>108281699
fuck yeah, glad to see Nitro+ XTX homies
what model you running right now

>>108281730
link?

>>108281737
it's not avaiable anymore, OpenAI nuked it

>>108281737
Dead, that's why we need local models.

File: 1762498714987760.jpg (291 KB, 1032x1536)

291 KB JPG

>>108281695
>>108281688

>>108281704
it's a five way tie between summer dragon, OG c.ai, mythomax, goliath, and midnight miqu

>>108281764
yet you don't use any of those, curious

>>108281771
when you "assume" it makes an "ass" of "u" and "me"

just woke up from my 12 hour coma
is qwen3.5 122b the new glm 4.5 air

>>108281794
Prove him wrong?

File: 47ywzmHFQp.png (356 KB, 1369x1324)

356 KB PNG

what you think? he pay or he pray?

File: pokemon.jpg (166 KB, 872x1344)

166 KB JPG

Tangential to /lmg/, but still pretty funny.

>>108281804
can you post pic i wanna try

>>108281804
cuckgpt

File: 1758966336880611.png (150 KB, 287x307)

150 KB PNG

>all this time later
>still no actual pixelspace, VAEless image edit model
>still no big, good omnimodal models that can generate images in chat
>still no big, good, natively multimodal models that "see" the image fully and properly
>still no real time voice conversation that you can have with the big, good models where they will also understand how you said something not just what you said
>still no basic real time 3d/2d avatars
>still no easy way to perfectly loop any image into an idle animation with ltx2/wan2.2
>still no good image 2 3d model
>ltx2 i2v still subpar
>even biggest models still get suck on things, still can hallucinate hard
>still no solved, just works, RAG
>still no solved, just works, internet search with something like searXNG
>still no just actually works browser usage
>MCP clients are still spotty, especially paired with spotty tool calling
>still no 1mil perfect context
>still no 3-10mil ok context
>still no infinite context
>still no 1T params 1b active SSDmaxxer model
and hundreds of more things

at least most big models are generally very good now and actually good enough to, with some help, vibecoood most actual projects you want
at least early moeGODS and ramCHADS won
at least z image turbo came out and was a huge leap in multiple big directions, basically solved resolution, almost solved out of the box realism (centered around portraits), huge speed boost
at least ltx2 came out and was a big turn towards faster genning, getting out of 5s hell, getting out of 720p hell, getting out of no audio hell
at least the great seedance 2.0 came out to be distilled by ltx3 or some other company this or next year
at least genie 3 showed that proper 3d space memory can be solved

everything can and will be solved but the lack of some more basic but important things like pixelspace image edit models or at least a basic 14-32b native speech2speech LLMs seems interesting.

>>108281813
tldr?

File: v7i0mvczmhoa1[1].png (245 KB, 345x614)

245 KB PNG

>>108281811
Got the pic from /v/, but I believe it's
>pic related

>>108281813
gpt 5.4 checks a few of those

I can run Qwen 27B at 1-1.5 token/s or Qwen 35B-A3B at 15 tokens/s.

>>108281829
gpt 5.4 doesnt exist

File: 1767422027240447.png (600 KB, 1186x1140)

600 KB PNG

>>108281825
>>108281804
werks on my machine i guess

>>108281688
https://www.stephendiehl.com/posts/computer_algebra_mcp/

when tf will they add mcp support to llama,cpp aaah. any program recs?

>>108281838
I imagine that there's a whole chat context we don't see that probably steered the model towards that sort of response.

Hello fellow anons. I need help with my qwen 3.5 27B Q5_K_M. its for some reason not thinking with each response its maybe 50% of the time and i have to retry the response to get it to think really annoying. im using koboldcpp btw is that the best backend? used previously ooga but it seems dead.

>>108281704
Me.

File: 1743998171556692.png (131 KB, 2048x286)

131 KB PNG

local sisters every time we start getting and edge the corpos fuck us in the ass, you are telling me they already have 5.4 sitting on a shelf?

Do people nowadays care if a model works with context-shifting or not?

>>108281877
>context-shifting
qrd

>>108281877
Yes.
When you send a bunch of requests to the model with just the last message changing, that shit is really useful.

>>108281879
Its a feature in llamacpp/koboldcpp that allows circumventing reprocessing of the whole context once you reach the max context you have set.

>>108281884
Qwen thinks otherwise it seems.

>>108281897
no i dont

>>108281902
are you Qwen?

>>108281897
You mean how llama.cpp can't do kv shifting with smm models?
That'll probably get fixed eventually.
Probably.
Eventually.

>>108281804
Every single time I read chatgpt's output I want to kys myself and do an hero.

>>108281907
No, that's a rnn issue, and it can't be fixed. if you remove a single token from the start you have to reprocess everything.

>>108281913
says the lobotomite

>>108281804
jfc what did they do to make it sound like this

File: 1748948632978330.png (507 KB, 946x1024)

507 KB PNG

Why are normies so dumb? And obviously the luddites are throwing a party not realizing this is a skill issue.

>>108281891
koboldcpp has that functionality under "fastforwarding", kobold's "context shift" purges old tokens from context when context is full.

>>108281926
You are too young to know what an hero even means. You are the real retard here.

>>108281936
either that is fake and gay or the company is fake and gay
either way it probably doesn't matter that the ai was also fake and gay

>>108281948
You are the newfriend, imagine saying you want to kys and an hero in the same sentence

>>108281936
That just goes to show that the company in question is worthless, that it doesn't really matter what they say or do, and that that their upper management is retarded and doesn't need to exist.

You are the reason why /g/ has died.

>>108281946
Doesn't work with rnn, still have to reprocess everything once you hit max context. Try running rwkv or qwen 3.5 and you will see that it won't work.

>>108281976
good

>>108281976
meant for >>108281928

I’m hearing good things about this “Qwen” model. Is it actually all that or can I go back to paypigging? I have 2x3090

>>108281978
Yeah I know, I'm talking about how it works in models where the feature is supported.

>>108281988
you need at least 2 6000s to run it properly, then it is legit better than opus 4.6

>>108281988
Try out either the 27B model or the 122B-A10B model. They seem to be roughly similar with the bigger model being a bit better and faster since it's moe.

>>108282000
Guess I’ll just fuck off then.

>>108282018
Yeah...

>>108281988
qwen2.5-72b fits on that at q4 which should be plenty

File: Screenshot_20260302_224736.png (359 KB, 1424x899)

359 KB PNG

Sillytavern/Kobold user. I may have altered a setting ages ago that I cannot remember, and now after every general prompt it just keeps going and gens another one after another after another. My token size per gen is 250. Surely there's something simple I'm neglecting here?

>>108282085
auto-swipes in ST user settings?

>>108281936
I don't understand how that happenes. If you feed the model your data and ask it questions it will have numbers to quote but if you don't give it any data why would you expect it to have access to your sales data.
Furthermore how do you not know your data well enough to do a sanity check simply by glancing at what it produces.

You have the same issues when you ask a subordinate to construct a report. You can't just assume he is correct and despite trusting him you must also verify the results.

I don't want to be mean but that guys issue is not AI.

>>108282094
thar she blows, cheers m8

I've been out of the loop for a bit. What's the current best local model available for utilizing large amounts of RAM with 32GB VRAM? Is it still DeepseekV3 and Kimi K2 or has something else been released?

>>108282099
why cant the ai figure out how to find and access the data on its own? isnt it intelligent?

File: 1mXpdOGQoj.png (290 KB, 1376x1233)

290 KB PNG

>>108282110
I still use this one

>>108282148
>isnt it intelligent?
No, stop falling for marketing lies like a retard.

>>108282155
i bet you are either very rich or very poor

>>108282165
so its smart enough to bomb iran but smart enough to figure out where the data is?

The bait will continue until anon's pattern recognition improves.

>>108282169
you wouldn't get it

File: 1748635088988770.jpg (374 KB, 2720x3000)

374 KB JPG

>>108281688

>>108281813
>still no 1T params 1b active SSDmaxxer model
You sleeping on snowflake arctic?

>>108282172
>so its smart enough to bomb iran
Sorting through communications in a network you already have backdoors in, doesn't require intelligence. An intern doing ctrl+f through the logs could have achieved the same result, albeit not as fast.

>>108282193
why she blushin

>>108282110
K2.5 thinking at q4

>>108282018
>>108281988
You don't need that. Your current hardware is sufficient to run Qwen3.5 122B-A10B or Qwen3.5 27B. Both are good models. If you want to do ERP with them though then you should grab the Heretic versions of those models.
>>108282040
This is an old model, don't use it.

>>108282203
q2 is better, more creativity

>>108282203
>>108282213
What are the gains and losses compared to K2-Instruct and K2-Thinking? Moonshot was hopping on the censorcuck train last I saw.

>>108282245
can you not use such vulgar words?

you crazy nigga. but i appreciate it.

File: nocap.jpg (400 KB, 1536x1536)

400 KB JPG

►Recent Highlights from the Previous Thread: >>108278008

--Agentic roleplay potential demonstrated through blackjack simulation:
>108278746 >108278774 >108278813 >108278819
--StepFun releases 3.5-Flash models and training tools:
>108280402 >108280421 >108280426
--122B model excels at Japanese text transcription:
>108278617 >108278679 >108279715 >108280042 >108280080
--Manual offloading outperforms --fit for 122B model on 3090+3060 setup:
>108281460 >108281492 >108281506 >108281543 >108281720
--International models lag behind frontier labs on ARC-AGI-2 benchmark:
>108279363 >108279384 >108279387 >108279404 >108279418 >108279428 >108279567 >108279598 >108279612 >108279657 >108279617 >108279629 >108279836 >108279469 >108279746 >108280473
--Open-source AI models performance gap with proprietary models:
>108279687 >108279804
--Qwen3.5-35B-A3B GGUF quantization benchmarks:
>108280652 >108280670 >108280678 >108280680 >108280735
--Qwen 3.5 Small Model Series release and performance claims:
>108278104 >108278328 >108280444
--Qwen3.5-35B-A3B-Heretic hitting 72 TPS on 7800X3D/7900 XTX with new llama.cpp:
>108281622 >108281636 >108281652 >108281657
--Qwen3.5 35b 4-bit vs 122b 6-bit speed tradeoffs:
>108280506 >108280525 >108280560
--Devstral-2 model's flawed Jinja date logic template:
>108278061 >108280633 >108280638
--AI response generation process critique and benchmarking culture:
>108278971 >108278991 >108279011 >108279036
--Qwen 3.5 benchmarks:
>108278349 >108278416
--AI internal reasoning resisting offensive prompt bypass attempts:
>108278112
--Qwen 3.5 27B speed optimization on budget hardware:
>108279596 >108279608 >108279623 >108279631 >108279638 >108279653 >108279662 >108279685 >108279689
--A.I. Dating Apps Complicate China's Efforts to Boost Birthrate:
>108278523
--Miku (free space):
>108278507 >108280771 >108281230

►Recent Highlight Posts from the Previous Thread: >>108278113

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

File: Base Image.png (1010 KB, 1080x3036)

1010 KB PNG

Multi-Head Low-Rank Attention
https://arxiv.org/abs/2603.02188
>Long-context inference in large language models is bottlenecked by Key--Value (KV) cache loading during the decoding stage, where the sequential nature of generation requires repeatedly transferring the KV cache from off-chip High-Bandwidth Memory (HBM) to on-chip Static Random-Access Memory (SRAM) at each step. While Multi-Head Latent Attention (MLA) significantly reduces the total KV cache size, it suffers from a sharding bottleneck during distributed decoding via Tensor Parallelism (TP). Since its single latent head cannot be partitioned, each device is forced to redundantly load the complete KV cache for every token, consuming excessive memory traffic and diminishing TP benefits like weight sharding. In this work, we propose Multi-Head Low-Rank Attention (MLRA), which enables partitionable latent states for efficient 4-way TP decoding. Extensive experiments show that MLRA achieves state-of-the-art perplexity and downstream task performance, while also delivering a 2.8 decoding speedup over MLA.
https://github.com/SongtaoLiu0823/MLRA
https://huggingface.co/Soughing/MLRA
neat. they tested at a 2.9B level so seems viable

>>108281622
>>108281652
Make sure you put RADV_PERFTEST=transfer_queue into /etc/environment too.

>>108282337
does this make ai as good as gpt?

>>108282375
luv me 7900 xtx

>>108281813
>>still no actual pixelspace, VAEless image edit model
stopped reading right there, you sound immensely retarded, next thing you know you be whining that we don't have bitnet yet

why don't we have bitnet yet

Is Qwen-3.5-9B better than Mistral Nemo for roleplay? I have very vanilla, normie tastes if that matters.

>>108282423
what that do

>>108281748
That right hand is fucked

>>108282443
racist benchod

Julius Caesar walks into a bar and says, "I'll have a Martinus." The bartender gives him a puzzled look and asks, "Don't you mean a Martini?" "Look," Caesar replies, "If I wanted a double, I'd have asked for it!" Another Roman walks in, holds up two fingers, and says, "Five beers, please."

>>108282427
No

>>108282427
If the Bijan Bowen video is to be believed all the Qwen small models are relatively good at the creative writing.

>>108282455
nta, but it's over...

>>108282451
The bartender is Julius' mother.

>>108282440
it doesn't lmao

>>108282451
Kek the first one takes me back to high school Latin class of which I remember little
>Semper ubi sub ubi
That is about it and that is not real Latin

>>108282455
Why not?
>muh erotica training data
Okay, but at what point does the raw intelligence of a model make that irrelevant. I seriously doubt that small, even technical, models don't have a single instance of the word "sex" or "horny" in them.
>>108282456
I guess I'll just have to masturbate to test it out then, huh. (I regret writing this but I'm posting it anyways)

Qwen small models excel in downstream tasks
I want the new AceStep, Wan, Anima, ZIT etc. to use new Qwen

>>108282456
I always believe all youtubers unquestioningly

>>108282464
If you want to use a model that is worse for your use case then go ahead, no one will stop you. You asked if it's better than Nemo for RP. It isn't.
>Why not?
Because qwen models are focused on math, coding and benchmaxxing. Creative work and general conversational abilities are an afterthought. It's also a smaller model than Nemo, which also works against it in terms of world knowledge.

>>108282480
Okay... have there been any new models at all that exceed the "creative writing" abilities of Nemo? I haven't been in these threads for about two months.

At this point I guess Nemo is almost 2 years old. Fuck.

>>108282460
>>108282463
ask your waifu to explain the joke to you

>>108282489
>have there been any new models at all that exceed the "creative writing" abilities of Nemo?
The problem with newer models is that more and more of their datasets are comprised of AI-generated data, leading to slop compounding generationally. There's plenty of better models for RP but a lot of people still prefer Nemo for its writing style, even if it is very stupid. Mistral Small 3.2 is a fair bit less dumb and similarly creative, but it's 24b. In the <24b range Nemo is still king. If you have a lot of RAM but only a bit of VRAM then you might look into GLM Air, I wouldn't necessarily say it's better than Nemo (though less dumb), but it's something different at least.

>>108282489
Not for ramlets

>>108282497
Got the first joke, but had to shamefully ask a model to explain the second.

>>108282509
the roman numeral. it's a good joke

>>108282505
ram is for poors, vram is for kings

>>108282497
The joke is obvious as it has to do with endings and that they denote singular and plural, I told you I took Latin and had to memorize that bullshit.
But you didn't say anythjng about my faux Latin underwear joke, sad

O

>>108282521
>ram is for poors
Not any more

>>108282504
What're the best options if Nemo's stupidity and lack of overall knowledge is too big of a dealbreaker? GLM Air was markedly worse than Kimi and Deepseek last I used it. Really it feels like Kimi and Deepseek are the only viable competitors and they're largely brute-forcing it through parameter differences.

>>108282522
Do you know what you feel if you dig your hands deep inside the ball area and go inside your body from the outside? That's how it feel to tardwrangle tat llm

>>108282528
usecase?

>>108282528
>they're largely brute-forcing it through parameter differences
knowing more is cheating

>>108282528
As I said, Mistral Small 3.2 is probably the only reasonable compromise for RP between Nemo and large MoE like GLM, DS, Kimi. It's far, far from perfect but there really isn't much competition. Gemma is too positivity slopped for RP, even if you get around its safety rails or use one of those stupid ablit/heretic tunes.

>>108282541
I think you are joking but it is funny how racist humans are against AI

>Oh it is only better at it because it studied more
How is that a bad thing.

>>108282522
probably 90% of the people here don't get the joke. that's like saying, "tengo un gato en mis pantalones"

>>108282534
Internally consistent fictional worldbuilding.
>>108282541
If it takes a model the size of Kimi or Dipsy to be marginally better than something a fraction of its size, we've either hit the point of diminishing returns on what the technology can produce or the underlying methodology needs refinement.

>>108282549
If GPUs with 128GB+ VRAM didn't cost more than most people's cars then no one would complain about models getting bigger.

File: zog.png (46 KB, 1168x517)

46 KB PNG

>>108282464
>raw intelligence
there is no such a thing. it's just a next token predictor. it appears intelligent in some situation because it has seen a lot of instruct and reasoning benchmax synthetic data that shows a simulation of a reasoning process about a variety of topics. It's still predicting the thing it saw in that data.
why do you think even the SOTA online API models will still behave like pic related when it sees any sentence related to their benchmax overfit? It doesn't have "intelligence". The entire purpose of a LLM is to take a document in the form of

<|some_magic_tag|>THE_LUSER
HERE'S A LOT OF RETARDED SHIT
<|some_magic_tag|>THE_ASSISTANT
HERE'S HOW I FIX YOUR RETARDED SHIT
STEP 1: KYS
STEP 2: INVENT A TIME MACHINE AND MAKE SURE YOUR MOTHER NEVER MEETS YOUR FATHER

and make document bigger. Until a stop token is predicted. Or get into an infinite loop and never stop until backend either timeouts or runs out of context like all GLM love to do.
MAKE. DOCUMENT. BIGGER.

>>108282557
If building a performant wasnt hard, there would be more options
I don’t want to defend leathernan, but with every bit of kit they release hitting msrp+50% or more instantly, they could easily charge more and increase profits for free rather than let scalpers and preorder lottery winners have a few bucks

>>108282559
>why do you think even the SOTA online API models will still behave like pic related when it sees any sentence related to their benchmax overfit?
Because safety and alignment layers are tantamount to performance sabotage and they're usually shoddily implemented by the brownest curry-stained hands at that.

>>108282570
I hear there’s an unaligned model at google they use internally for strategy

File: 1753598247563160.jpg (75 KB, 847x477)

75 KB JPG

>>108282569
There isn't anything special about nvidia GPUs though, CUDA could easily be replaced by Vulkan/Rocm with similar performance but AMD is controlled opposition and Intel are still recovering from a decade of being complacent jews doing nothing
The margins on nvidia cards are already through the roof, they could double the VRAM of the 5090, sell it for half the MSRP and they'd still be making a decent profit per unit.
>[COMPANY] isn't (yet) fucking you as hard as they could be (though their thrusting is still getting harder every year)
gee thanks

>>108282557
That's only an issue because of human greed, at any time there are hunedreds of thousands if not millions of vram gbs doing nothing that we could be using

>>108282586
Likely true. Anthropic very likely has an unaligned Claude version as well and revoking access to it was likely the cause of being labeled a supply chain risk by the Pentagon.

File: 1337941191745398.jpg (35 KB, 451x445)

35 KB JPG

>>108282514
the wha--- ooohhh

>>108281877
Uhhh yes?
Have fun prompt processing that whole ass context again for every little shit.