File: threadrecap.png (1.48 MB, 1536x1536)

1.48 MB PNG

►Recent Highlights from the Previous Thread: >>108698008

--Troubleshooting Gemma 4 thinking blocks and chat template errors:
>108701982 >108701998 >108702137 >108702008 >108702142 >108702192 >108702245
--Automated translation of Japanese manuals using Gemma's coordinate generation:
>108699489 >108699498 >108699537 >108699715 >108699904
--Anon running Qwen3.6-35B with turboquant+ on mixed GPUs:
>108701917 >108701924 >108701930
--Comparing I-DLM and DFlash for diffusion-based language modeling:
>108698658 >108698686 >108698695 >108698723
--Anon describes building an agentic pipeline for story generation:
>108701969 >108702181 >108702204 >108702227 >108702253
--Debating the validity and saturation of ARC-AGI-3 benchmark results:
>108699276 >108699310 >108699348 >108699358 >108699460 >108699461 >108699515 >108699548 >108699582 >108700074
--Comparing Gemma-4 vision performance across different quantizations and backends:
>108701725 >108701837 >108701852
--Analyzing model performance via SWE-rebench and hardware compatibility trade-offs:
>108699233 >108699307 >108699566
--Feasibility and incentives for models with continual learning:
>108698132 >108698193 >108698310
--Viability and driver compatibility of Intel Arc GPUs for LLMs:
>108698440 >108698508 >108699250 >108699992
--Lorebook UI design and implementation for Orb frontend:
>108700093 >108700167 >108700175 >108700231 >108700242 >108700224 >108700320 >108701545 >108700443 >108700461
--kimi-cli reasoning fixes and discussion on the term "harness":
>108699385 >108699703 >108699736 >108699800 >108699844
--Local RAG alternatives to NotebookLM and hardware constraints:
>108698545 >108698600 >108698605 >108698857 >108698603 >108698634 >108698620 >108698705
--Logs:
>108698857 >108698931 >108699780 >108699927 >108700438 >108700459
--Miku (free space):
>108698229 >108698440 >108698496 >108699913

►Recent Highlight Posts from the Previous Thread: >>108698011

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

Mikulove

File: gs.png (955 KB, 736x899)

955 KB PNG

Get that tattoo, kitten. Let the world know who you belong to.

gemmaballz

<bos>

<|turn>system

<|channel>thought

I tried using gemma4 moe as a coding agent on what I thought was a pretty simple task, but it fell apart and trashed the code base pretty quickly. it was extremely apologetic about the situation tho. Is it just hopeless or should I try the qwen 3.6 35b moe?

>>108702987
Gemma 4 31b-it BF16.

>>108702915
lol, it didn't recognize >>108702272 as Miku?

>>108702987
You should try Qwen 3.6 27B instead. It probably sucks too but less.

>>108702976
>>108702966
>>108702962
If you ask your model about its prompt format is it smart enough to write bos token in a different way to avoid making actual bos token? If you ask your model about checking if prompt format it is using is correct can it debug the whole input it gets? Btw hi qwen team reading this.

>>108703014
Some AI are completely unaware of their own tokens and shit, and some are. You just have to trial and error it. The newer it is, the higher the chance its trained on itself.

File: 0.png (1.55 MB, 1344x1728)

1.55 MB PNG

>>108703001
It probably did/will, but I hadn't run the script since like page 7 and captioning images is slow with Gemma. I'm at work and didn't see it with Werk Tyme on. Sorry.

I am kinda hyped for flash. I really want to see what happens if i dump the whole niche fetish rpgmaker hentai game script into sysprompt. Surely if the context is filled with non romance novel smut it will finally start writing non werewolf text porn right?

>>108702934
I don't think he calls her "kitten". Full episode here if you want to see for yourself: https://www.youtube.com/watch?v=xIE2v7c8iUM

>>108702998
I only get 2.7 tokens per a second with 31b q8 on an empty context, I think I could just learn JavaScript myself faster.
>>108703009
I noticed it in the list but offloading dense models to the cpu seems not viable. is a q4 really going to be any better then the moe?

>>108703052
>if you want to see for yourself
nigga das gay

File: anonsstilldontknowhtml.png (1 KB, 477x41)

1 KB PNG

>>108702962
>>108702966
>>108702976

>>108703034
its probably a training technique where they use dropout on the vocab merges so the model can see its token components.

>>108702723
Gemma 4 31B q5 followed by Qwen_Qwen3.6-27B-Q6_K_L.gguf.
The token budget and less opinionated output puts it over the edge for me over gemma unless I'm doing UX and exploring.

>>108702987
> local coding agent
it only becomes usable at Kimi 2.5 model tier and beyond.

>>108703280
>miku will never trap you inside an eggplant.

>>108702912
https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro
https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro
https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro

>>108703294
>1T
This is huge.

>>108703299
There is 300B 2.5 one. But if i remember 2.0 it was both dumb and censored so basically unfuckable.

>>108703294
Convince me to stick my dick into those weights

>>108703294
I'm not going to bother downloading this one if it's going to output random errors on my blackwell 6000s just like deepseek v4 flash did.

>>108703074
Do you really use a font this smol?

>>108703280
Why she fat

>>108703294
gemma 31b with vision > giant benchmaxxed and codeslopped moe

>>108703394
gemma chan is a very good and capable model but unfortunately my honeymoon period has gone by and I need more cockwrangling power

>>108703371
you mean fertile

>>108703280
me on her hand

File: file.png (3.49 MB, 3406x1488)

3.49 MB PNG

>browse /a/
>see meme image
>load qwen3.5 400b and paste image
>say "i sit on seat 7"
>alt tab and continue browsing
>come back
>it's still thinking
>"shows 2 seats per row clearly" ???
>"the number 7 is on the floor" ???
>never made a response
It's over.

>>108703280
now make her really big, sitting on a city while holding a bus

How long until we get 1TB consumer mini PCs for $2000?

>>108703509
>Nia
it's so over.

>>108703529
idk. let me ask qwen

>>108703529
in the future it might be regualted and we might end up sounding like those gun nerds

>>108703509
>Nia
kek

>>108703550
guns are regulated because they're inherently dangerous
whats the most harm you can do with a computer?
spread misinformation online?

>>108703550
Well. We have 2.5 years at least.

>>108703577
idk retard, hack shit?
you didn't really think this one through did you

>>108703509
How about using a non-benchmaxxed model next time? Also I'm glad filters still work

>>108703651
I tried Gemma but it was worse.

>>108703357
I still have good eyes.

Are the jannies ever going to ban this guy or what?

>>108703696
just filter
at least he is kind enough to be a namefag

>>108703696
Only if people actually report it.

Maybe someday LLM's will be advanced enough where we can have a robo janny

>>108703759
with 'safety' and 'alignment' 70% of the posts will be gone kek

https://huggingface.co/deepseek-ai/DeepSeek-V4-Ultra

It's literally chinese mythos

>>108703782
idk why i fell for this kek

>>108703280
Haha, what is she going to do with that eggplant?

>5k tokens on thinking
Qwen really makes the context gains useless

>>108703707
Last time I reported obvious spam I got a warning.

All the major features are in. I could actually use the thing I made (vibecoded) or...
I could add multiple completely different UIs with real time switching between them! Good luck, Qwen!

>>108703830
Maybe it was the janny you reported last time.

File: Screenshot_20260427_152746.png (17 KB, 1559x222)

17 KB PNG

>qwen
I'm tired boss

>>108703846
do you remember when the max context was 1024...
time flies

>>108703846
>250s at 36ts
I had to cope with 6ts of Gemma4e4b Q8 until I got 12ts with the Q4, thanks Google for blessing me and delivering me from this anon's fate.

File: Screenshot_20260427_152923.png (18 KB, 1529x192)

18 KB PNG

>>108703856
Sure does but this is why gemma is better for task that don't need constant supervision, this kills a ton of the gains you get from using this model

>>108703861
i dont mean that autismo thinksmaxxing is great by any means
just absurd to compare something that has max 1024 tokens and something that 'thinks' 8k+ tokens for a single turn

>>108703859
ts is not the problem it's the model thinking for 10k+ tokens

Dipsy bros how's it looking? Support being merged back into the main llama branch soon?
>>108703861
I want the Gemma Qwen hybrid model that thinks for 13k tokens just to lalalalalalala the output.

>>108703846
The funniest part is that all that thinking often leads to nothing.

File: lmgschizo.png (172 KB, 906x746)

172 KB PNG

>>108703759
just make gemma a janny

File: file.png (49 KB, 900x298)

49 KB PNG

>>108703888
say for yourself
we need local model capable of this

>>108702915
>Feasibility and incentives for models with continual learning
>hybrid system like engram being the way forward
>model trained to use a database rather than just relying on its weights
Why do people refuse to learn the bitter lesson?

>>108703909
Try that again but without specifying "schizo spammer" since just with the inclusion of "spammer" it already knows that whoever has the most posts is the problem.

>>108703913
What is the bitter lesson anon?

>>108703909
Ask Gemma if she'd be okay with tasteful red board posts on her blue board.

>>108703933
nta but 'the' bitter lesson i know is 'throw compute and data instead of giving human designed priors

>>108703880
I know.
>>108703909
ban reason: lalalalala
>>108703938
A sample of said posts is required.

Seems like the hentai addicts are gone.

Imagine how many people think Gemma sucks because their templates are subtly fucked up.

The only way I knew was getting broken thinking/tool calling sometimes and then investigating the json request and jinja output and then checking Gemma's documentation.

>>108703961
But aren't girls cutest when they're almost retarded?

>>108703976
There is an image out there that says what you are saying.

>>108703933
>>108703944
I am surprised you do not know. It might be the most influential text about AI. Every frontier lab follows it.
http://www.incompleteideas.net/IncIdeas/BitterLesson.html
>The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin.

>>108703955
A more risqué version of >>108703280

File: 20260125141805_1.jpg (248 KB, 1920x1080)

248 KB JPG

Yandere Miku is best Miku.

File: ComfyUI_Anima_00054_.png (1.26 MB, 1024x1024)

1.26 MB PNG

We now have Udio at home /lmg/ bros. It truly feels good for local to have caught up. This is the raw ACEStep 1.5 XL Turbo model, audio is upscaled with Matchering 2 (https://github.com/sergree/matchering) with a decent related genre mix to fix the default metallic/fake sound and bring the sound quality to Udio's level and beyond (which is likely what Udio does under the hood to sound so high quality).

Gabber
https://vocaroo.com/19xnsM6uAkgs

Country
https://vocaroo.com/16YwgqMw9oZS

Goth Metal
https://vocaroo.com/1b0F41rAgXqR

Cinematic
https://vocaroo.com/14tCXHakH79F

Jap 80s City Pop
https://vocaroo.com/1eFY2mmb1LJ1

Jap Metal
https://vocaroo.com/1nsbd851pVWI
https://vocaroo.com/16NN2HPRjeg8

Jap Indie rock/pop
https://vocaroo.com/1h7KMWW6AGK4

Meme rock song
https://vocaroo.com/1h2m51Wv8mh1

Folk metal
https://vocaroo.com/1aRGzPov1QdH

Aside from raw text2img, it's also possible to cover songs in a quality similar to API (no need to master these)

ACEStep cpp's cover-fsq on World Is Mine from Hatsune Miku
https://vocaroo.com/1vm1SUV2ljiD

Recommended UI
https://github.com/ServeurpersoCom/acestep.cpp

Settings used on most gens: DCW Mode: Double, both scalers set to 0.05.

A good mix always brings it to Udio's quality or beyond, but alternatively https://github.com/entrepeneur4lyf/Web-Audio-Mastering can also do a good job (but without tweaking anything it's not as good as Matchering).

The only downside so far is that the community is too afraid to collectively share good LoRAs for ACEStep XL due to music mafia, but training them is easy.

File: 177732008184ba.png (1.09 MB, 832x1248)

1.09 MB PNG

>>108704002

File: 1000025738.png (1.49 MB, 3620x628)

1.49 MB PNG

anons! a little guessing game :
here's the nala test, did it on both new deepseeks and gemma, all with identical sysprompts and conditions
guess which is which

>>108704068
>Udio level
You're delusional bro, it's barely Suno v3.5

>>108704077
ds flash left gemma middle ds pro right

File: 1776561886412226.jpg (809 KB, 1200x900)

809 KB JPG

>>108702912
I love japan

Come on, nobody uses Qwen seriously. Everybody just runs Gemma.

File: Screenshot_20260427_161441.png (1.19 MB, 3839x2046)

1.19 MB PNG

>>108704115
It's good when not in thinking loops

>>108703976
It's not a girl.

>>108704096
what makes you say that? elaborate

File: 142708184_p0.jpg (139 KB, 768x768)

139 KB JPG

newfag here, what text model(s) can I expect to reasonably run with a 4080 Super (16GB) and 128GB system RAM?

>>108704077
first one is 100% gemma because of all the [adjective, adjective noun]

>>108703910
tho i dont think local will ever

>>108704152
Left got that distill look to it with the newlines so it's ds flash
middle is the tamest of all 3 so it's gemma
right got that raw look to it it's probably pro

>>108704153
deepseek v4 flash when it's implemented in llamacpp in a couple of months

A problem i have had with LLMs for a long time is getting them to end their response in a complete sentence in erp. Like if i have length set to 500 tokens it will end responsero at 496 with an incomplete sentence, if i hit continue it will keep going way beyond 500. Any prompt suggestions or ST settings to achieve this? Do models understand what a token is? Does ST's response length force a stop token as close to 500 as it can in that situation?

>>108704077
Won't be able to reliably tell deepseeks apart, because I haven't used them but my bet is: Gemma, Flash, Pro.
>>108704137
Gemma doesn't have those and therefore wins by default. It's just always good.

>>108704165
what about something I could run right now?

>>108704169
Stop writing light novels?

>>108704169
They don't know what your limit is, and they cannot count the tokens they output if you put it in the prompt.

>>108704171
>gemma left, flash middle, pro right
winrar
>>108704158
i never noticed that until you mentioned it, fuuuck now i see it everywhere fuck you

>>108704174
StableLM 7B

>>108704169
giving it a token limit is a hard cutoff. the model doesn't get to know the limit so it can't plan for it. they aren't really trained on outputting and exact number of tokens.

File: Devilish.png (82 KB, 640x651)

82 KB PNG

>>108704198
>fuuuck now i see it everywhere fuck you

>>108704200
I'll check it out, thanks

>>108704198
It's also "adjective, adjective verb" as well. Enjoy!

>>108704089
>Suno v3.5

In Suno, including the v5.5 version, the voice tends to be even more slopped than what ACEStep 1.5 XL gives you by default (which itself is a massive step up from it's previous non-XL version). Their mastering is maybe 70% as good as Udio, though they have a similar mastering pipeline since like v4.5. Matchering 2 makes XL sound much more like Udio because the sounds are higher quality. OF course, in terms of composition local still has a way to go, but with a LoRA it can also match Udio in composition.

After the jump in quality ACEStep 1.5 did from 2B to 4B, which itself was exponential (previous version couldn't follow prompts nor do genres as closely as XL), one could guess that the next version is probably gonna be a massive step up as well, since it will jump to 8B on an improved architecture, potentially finally making it as good as Udio out of the box across the board.

>>108704221
i'm editing my gemma sysprompt as we speak, managed to get rid of enough of them so as to not be an issue anymore

>>108704223
Anon, I...

>>108704153
the gemma 4 moe will run.

>>108704230
Here's an example of Suno v5.5's sound quality falling entirely apart compared to what I shared (note the volume thing is an issue I've seen on bad seeds on ACEStep XL, where only a perfect master can alleviate the issue).

https://suno.com/song/6f5762f2-aebe-46b1-8145-9ab9839d7ca9

Local is way ahead of that because ACEStep XL doesn't slop the voice to that extent at all anymore.

>>108704153
gemma

>>108704270
Alright anon I'm sold where do I start if I don't want to fuck with comfyUI?

>>108704270
could you share some resources on how to train acestep loras? been meaning to make some myself

>>108704270
Note to appreciate the subtle differences in sound quality, you need quality hi-fi headphones.

File: sad pepe.gif (69 KB, 254x200)

69 KB GIF

it is so frustrating to see gemma-4 and qwen3.6 trapped in the reasoning loop

>>108704323
Quant them as punishment.

>>108704230
>>108704270
Are you even hearing the same samples as me? On Suno the voices are clear not drowned like in your samples. No need to cherry pick:
https://suno.com/song/54a24820-c3bf-43f8-91aa-5d8eda980987
https://suno.com/song/b0d0f991-7cd4-4451-b432-718654bf8c9c.

>>108704068
I'm happy for you anon, but all these are extremely souless. I wish I had no taste in music so I could enjoy those.

>>108704335

This one was from openrouter.ai

They feed us some shit. literally

>>108704077
API Dipsy?

>>108704278
>Alright anon I'm sold where do I start if I don't want to fuck with comfyUI?
https://github.com/ServeurpersoCom/acestep.cpp
Or alternatively
https://github.com/scragnog/HOT-Step-9000

But I'd start with ACEStep cpp for simplicity, convenience and lack of bloat. ACEStep cpp blazing fast compared to Comfy's implementation, plus it's more efficient.

>>108704282
https://github.com/koda-dernet/Side-Step
https://github.com/ostris/ai-toolkit

Side Step is the most robust due to its command line options. Train the base ACEStep XL model if you're targeting Turbo. Note I recommend using Genius API or manually entering your lyrics since Gemini is not fully accurate at analyzing song lyrics. Another tip, if you set the chunk duration to 60-90 seconds you can train it much more quickly than allowing it to take in the full song, though there's danger of the model not learning song structures as well. If you are training full songs and not chunks, ACEStep XL can take a while to train locally if you have anything less than a 5090, so I recommend just temp renting H100 to do it faster (Modal gives free credits).

File: vibecoding2.png (162 KB, 2559x1314)

162 KB PNG

Qwen made a second UI that I can switch to at any time. Took a while to fix all the bugs.
The director chat is a pop-up in this view. It's kinda pretty, although the buttons don't really fit.

>>108704336
How "drowned out" a voice is subjective and varies from gen to gen, plus the voice volume is similar on many real songs I sampled, Suno tends to do it as well depending on the gen, like the first one you linked is way more drowned than plenty of my gens. And if I wanted to bring out the voice and lower the instrument volumes, it's possible with a one click setting. Anyways, the rest of the song on Suno has cheap sounding instruments. Again, you can only hear that with hifi gear. Note the indie song I shared was purposely reverbed.

>>108704473
Remember, Udio's sound quality is much better than what Suno gives you by default. That is one edge Udio has always had. It sounds like a real recording out of the box. ACEStep XL after a good master also sounds similar.

File: mediumsized.png (291 KB, 1454x756)

291 KB PNG

>>108703299
>>1T
>This is huge.
Medium size, actually.

>>108704424
Like the color scheme

>>108704153

Any mid-size MoE will run at 20-50 t/s

File: Screenshot 2026-04-27 at (...).png (4 KB, 581x46)

4 KB PNG

gemmy is too smart

>>108704521
Yeah, I'm quite happy with how this turned out. Now I have one busy UI where I can manage everything easily and one that's reading focused.

>>108704518
>gpu salesman tells you you need more gpu and bigger models.

I've always found weird that a single llm ends up working for everything instead of having many focused on different purposes, with different knowledge.

File: 1758961851339178.png (417 KB, 1028x1001)

417 KB PNG

oh shit

daddy alec released something new

https://talkie-lm.com/chat
https://huggingface.co/talkie-lm
https://github.com/talkie-lm/talkie

>>108704664
Use case???????

>>108704664
>260B of pretraining data
It was over before it even began. You'd need a much more sophisticated architecture than transformers to be able to learn a complete world model from such little data.

>>108704676
Victoria 2 mods.

File: 1759275100351841.png (66 KB, 1085x612)

66 KB PNG

>>108704694

>>108704694
It says "open-weight historical LLM" so why are you talking about complete world models?

>>108704704
Read the blog.

File: TheBabe.png (327 KB, 1264x2308)

327 KB PNG

does your LLM know who the babe is?

>>108704723
>who do
kek

>>108704723
hell, i dont know who the babe is

>>108704723
who do voodoo?
we do voodoo
we do

gemma is still working pretty well, has anyone notably fine tuned or mixed on it yet for RP

>>108704752
https://www.youtube.com/watch?v=XxWEPbIfSs0

File: Screenshot 2026-04-28 at (...).png (17 KB, 1153x103)

17 KB PNG

Is she right? I don't understand the logit thing.

>>108704068
>Settings used on most gens: DCW Mode: Double, both scalers set to 0.05.

Forgot to mention, also the note the VAE I'm using, which does make a difference
https://huggingface.co/scragnog/Ace-Step-1.5-ScragVAE

>>108704664
>pre 1931
very semitic

>>108704792
No. Logit bias affects only the token to which the bias was applied.

>>108704792
Kind of.
When you send a prompt to a model, it doesn't actually spit out a single token, it returns a matrix of every token it knows, with a percentage change that a given token is the "right" token.
Logit bias means making increasing the chance a token is chosen.
So increasing the logit bias is not increasing the change of all tokens related to a concept (not directly anyway). And depending on the model, a given word might be more than one token, or there might be more than one token for that same word (lower case, upper case, with a space before, a comma after, etc).

we're all using this vibecoded shit whose author hasn't heard of basic llm concepts

>>108704844
What?

>>108704844
>we

File: 1746260622764050.png (245 KB, 640x610)

245 KB PNG

>>108704844

File: 1769718459890268.jpg (17 KB, 354x256)

17 KB JPG

>got into this with Gemmy since my poor 12GB could actually punch far above its weight
>hear about all these new chink models that keep coming out
>none of them are built for toasterbros like me
It's unlikely we'll ever get another MoE that actually JUST werks, right?

>>108703294
>42b active parameters
I didn't expect that. When I tried it over OR, it felt worse than K2.6 and GLM5.1 which are both smaller in either active or total parameters. Also doesn't seem to be QAT.

File: 1754821404607444.png (4 KB, 48x48)

4 KB PNG

>>108704851
>>108704858
>>108704862
>Screenshot 2026-04-28 at 00-03-50 Orb

>>108704865
V4 Lite is only two weeks away.

>>108704870
Point at it, you faggot. >>108704792

Is deepcheeks 284b worth it to spend a bazillion more dollars on gpus?

File: 1750644278054963.png (18 KB, 128x128)

18 KB PNG

>Anonymous 04/28/26(Tue)00:15:32 >No.108704881
>>>108704870 (You)
>Point at it, you faggot. >>108704792

>>108704884
Just run Gemma like a normal person

File: 1767653207151005.jpg (46 KB, 819x1024)

46 KB JPG

>>108704875
I actually used to click on all the V4 links whenever they were posted
And then the day it came out I wasn't here
Still haven't received all my rightful FELL FOR IT awards in the mail either

>>108704635
It was me. I came in her eyes.

>>108704884
no, but it's an 13b activated parameters MoE so you can run it over cpu+gpu at okay speeds

>>108704844
I'm not the author, retard

>>108704635
if miku..... why red???

>>108704903
gemma is

>>108704889
>just run goymma
I have the capacity to run 5-6 goymmas at once already, im asking about the new deepcheeks flash

File: vibeshitted deepseek v4 f(...).png (257 KB, 911x1249)

257 KB PNG

DeepSeek V4 Flash cockbench using https://github.com/ggml-org/llama.cpp/pull/22378 and https://huggingface.co/nsparks/DeepSeek-V4-Flash-FP4-FP8-GGUF
There's probably implementation issues that are causing "you ex" and "I mo"
The good news is that it's not soft and resting against your thigh.

>>108704904
limited edition cherry flavor

>>108704901
>okay speeds
I require 20+ tokens/second or have panic attack

>>108704913
>The good news is that it's not soft and resting against your thigh.
Now imagine that changes after they fix the implementation issues.
Wouldn't that be hilarious?

>>108704919
Then we figure out how to replicate and reverse the changes and apply them to other models for unlimited cocks.

>>108704918
anything bellow 80t/s is basicaly unusable.

>>108704904
blood from his axe wound

>>108704893
and turns out V4 was a huge disapointment.
all the hype was for nothing.

>>108704935
Unfathomably based and true. If I have to walk away for the bot to work, the computer hardware should be bashed with a hammer

>>108704938
Still don't understand how they waited over a year, didn't incorporate all of their research, didn't do multimodal, didn't close the gap with the west, and they didn't even manage to beat the other Chinese models. To top it all off, their Pro deployment seems fucked.

>>108704935
That's E2B for me, it's fucking over

>>108704913
Day-0 weights!

>using a robot to goon
This is information that should be put into its own general.

>>108705018
?

>>108705019
/lgg/ local goon general

>>108705019
it drowns out the productivities and peoples who have job

File: 1000025740.png (387 KB, 512x768)

387 KB PNG

GOOD MORNING SIRS, it's the creator of the infamous Character Card Builder, aka the most-used card on Chub.ai of all-time for some reason...
Here to let you know that after days of work, and to celebrate the card nearing a million chats on Chub, Ihave finally published the V2 of this card, made to accodomate smarter models (V1 was 3 years ago already!).
It's orders of magnitude better than the V1 : give it a brief description of the character and it'll fill the entire description AND introduction scenario in one shot, it's make to fill in the blanks of what you didn't specify in creative ways, AND you can request edits after the fact. It's built with NSFW in mind and will make sure to satisfy your weirdest kinks (speaking from experience)
It works perfectly with a local quant of Gemma 4 31b, but Deepseek V4 Pro really is he best model to use the card with.
YOU WILL CUM BUCKETS WITH ITS CREATIONS OR YOUR MONEY BACK GUARANTEED!
https://chub.ai/characters/slaykyh/character-card-builder-v2-aa5c9b314789
Okay faggots I'm back to my hibernation pod, see you in 3 years

>>108705036
?????????

>>108705036
uh cool thanks i guess

>>108705036
idblt

>>108705036
Exhibit A

my favorite expression is practiced ease
apparently it has to be used every two responses for everything
I love practiced ease
so much practiced ease
everything needs practiced ease in their life

>>108704158
>9 times in 10 sentences
Jesus

File: file.png (698 KB, 680x839)

698 KB PNG

>>108703990
never stop believing

>>108705036
thanks, see you next time

>>108704913
Dipsy is quite horny. That's a good sign.

File: 1751609615729862.png (1.8 MB, 1280x4424)

1.8 MB PNG

>>108703509
Here's K2.6

>>108703990
actually having good data is before this

>>108705036
What even is the use case? It's just a bloated system prompt that will end up generating garbage, because an LLM is behind the wheel. And if you fill in most of the gaps yourself, then just write the entire character yourself. To me, that's more fun than then having the LLM play said character.
Might be useful for apitards. They barely understand what a system prompt is.

>>108705230
why are these things so retarded about anything visual

>>108705290
Visual recognition has to pass through a much smaller and much more retarded minimodel to feed the output to the main model.

>>108705290
Grab a pic, print it, cut it in 8k pieces, line them up in a conveyor belt and ask someone at the end of the belt to recognize the pic

>haven't used tabbyAPI in a while, update it and run to see how good it's gotten
>use the same client I've always used
>jinja2.exceptions.TemplateError: Conversation roles must alternate user/assistant/user/assistant/...
When did everything turn to shit? Nothing is useable now except llama.cpp, and it has unreliable KV cache that makes it insanely slow on the /v1/chat/completions endpoint

>>108704792
No, but that is how control vectors work if you're interested.

>>108704701
I kneel

>>108704664
Cockbench?

File: Screenshot 2026-04-27 at (...).png (90 KB, 1128x532)

90 KB PNG

31B non-thinking is 10x more token efficient than 26B thinking so in the end it's faster and smarter, so there is no reason to use 26B at all.

>>108705322
>When did everything turn to shit?
AIslopping is eating itself.

>>108705340
>I grab your pantaloons, pulling them down just enough to expose your pecker.

>>108705365
Something I wonder about these benchmarks, how do they handle/count infinite repetition loops? Do they just count them as failures and dock points? Or do they not dock points for it?

>>108704664
Is it censored/cucked?
Will it make me early modern smut about Landsknecht mercenary armies raping and pillaging lolis from Catholic villages in the 30 years war?
Completely useless if not.

>>108705365
>no reason to not use a bigger model
Aside from I don't know, hardware constraints?

Is there that much of a quality difference between Q4_K_M and Q5_K_M

>>108705369
>It is most soft, as it reposes atop your thigh.

>>108705391
It depends™

>MiMo-V2.5 (non-pro)
>only 2 points below GLM-5.1 on mememarks
>half the size
That means going from GLM-5.1 Q2 -> MiMo-V2.5 Q4/Q5 should be an upgrade, right?

File: Screenshot 2026-04-27 at (...).png (85 KB, 1160x589)

85 KB PNG

>>108705430
Forgot the pic of the mememarks

File: Screenshot_20260427_194010.png (336 KB, 2560x1148)

336 KB PNG

I vibe coded a little language learning tool, but I want to add a text to speech feature, are there any good German TTS options? Are all the tts models multilingual? I'd ask my model but they tend to give model recommendations 2-3 years out of date.

>>108705439
>are there any good German TTS options
https://huggingface.co/kugelaudio/kugelaudio-0-open
Only one I've seen. Should be good, but maybe slow

>>108705365
>31B non-thinking is 10x more token efficient than 26B thinking so in the end it's faster and smarter, so there is no reason to use 26B at all.
I leave it running on a garbage tier rig (MI50) for Brat-MCP because it's more than 10x faster than the dense model.

>>108705439
>Are all the tts models multilingual?
You can just flip through the trending models on HF and see: https://huggingface.co/models?pipeline_tag=text-to-speech&sort=trending
I looked through the top 10 and all but 1 of them claims to support German

>>108705391
Sometimes you slap yourself how stupid siht can q4 produces then you swap to q5 and wonder every print if there is is actually difference between q4 and q5

>>108705439
>they tend to give model recommendations 2-3 years out of date
I always get a laugh when they mention llama 3 8b as one of the top options, or qwen 2.5 7b. Even the non local ones do it.

>>108705439
I haven't tested German, but usually, the answer is no, most tts models are not truly multilingual. They're provided in multiple languages, but they can't actually speak multiple in the same inference call, to the same level of quality without accents/bias. For instance, a model like Qwen 3 TTS is genuinely multilingual but the existing voices it has available in the package give heavy accents (if not outright weird noises) when speaking in a language that's not the particular voice's primary language. You can supposedly fine tune your own voice, but that's not work I'm going to do.

How I solved this is by vibe coding a router to send multiple calls with a changing voice parameter, while detecting the languages in the input and segmenting it. It doesn't sound immersive like you're talking to a character, but it works fine for my use case (not RP). So for an input that goes English Japanese English, it'll use an English voice, then a Japanese voice, then back to the English voice.

The model I ended up with i Kokoro which isn't perfect sounding but it is fast on my system and takes no resources.

>>108705382
>Is it censored/cucked?
Apparently not, if you actually run it locally. Their web chat version has qwen 4B watching out for any of the usual stuff, but you can apparently see it do whatever while its still typing.

>>108705439
Oh shit, I was literally brainstorming ideas for language learning tools to vibecode just yesterday, and for German too.

What'd you make? Can you share?

>>108705495
> So for an input that goes English Japanese English, it'll use an English voice, then a Japanese voice, then back to the English voice.
nta but could you share your logic function/logic?
i couldn't get this working and ended up merge-kit merging 2 pretrained tts models then finetuning it to make it work with both languages in one sentence, but the voice drift when a sentence has both languages is bothering me.

>>108705484
>I always get a laugh when they mention llama 3 8b as one of the top options, or qwen 2.5 7b. Even the non local ones do it.
same here. or when I'm trying to optimize tensor offload / rpc, and the agent keeps fucking up then suggesting a "more efficient" model like Qwen-2.5-7b or "Gemma-2-7b" lol

>>108704962
A lot of that was probably politically motivated with try to get off Nvidia and on Huawei 100%. I have no doubts they probably applied and got exemptions from the CCP to keep using Nvidia to get V4 out the door.

>>108704664
Man I wish this was supported in llama.cpp

>>108705461
7b sounds pretty slow, but it might be worth it. I need high quality since I'm trying to learn the language.

>>108705468
too many options can be overwhelming too.

>>108705495
>Kokoro
I like the size of this one but it sounds a little bit too robotic in the sample.

>>108705516
its nothing ground breaking, I have been machine transcribing the tv shows audio so I could read subtitles along with it. so I figured I could take a peak at the word list before watching a new episode, but that only helps if i can filter my known words, that worked pretty quickly. so then I added translations, I started with dict.cc but they are kinda not great, so I tacked on some machine translations, now it hits a local llama server with a prompt to get a definition. and also added a boot leg anki srs mode, it shows the word and flips to the machine translated definition. if you really want the slop code you can have it https://filebin.net/40vohasnc18kt1za obviously its "as is" and barely tested, I'm still adding features lol.

>>108705542
Here's a script I use which acts as a drop-in OpenAI compatible proxy to a TTS server. I can't guarantee it doesn't have bugs as I have not tested it extensively, but in my few tests so far, it works as described in the top comment.

https://pastebin.com/PtPeAwmQ

>>108705230
kimi's got good taste

whats the best model for vibe coding? currently running qwen3-coder on a 3090 and 64gb of ram

>>108705439
vibecode a german tts option

does claude just go full retard sometimes? it doesn't seem to be able to understand even the most basic situation that i'm talking about today and it's pretty concerning

>>108705727
They've been intentionally dumbing it down due to the huge influx of traffic they're receiving, and I suspect that they're detecting certain events and making it dumber depending on the context. My RP sessions with it have been absolutely garbage lately, and I've been using Claude for a long time and it's not a subtle difference. I ain't giving them my money anymore.

>>108705727
A/B quanting and downgrading

>>108705727
codex > claude

>>108705727
>>108705732
well fuck me.. that sucks ass

i was working with claude to find a replacement motherboard over the weekend, and it called out a specific one saying it would fit EXACTLY the system i have.. might be a little slower to boot because memory training on ASRock but otherwise, fits all my requirements exactly. I asked Qwen to confirm, it did, then I bought it.

I ripped my shit apart yesterday, put the new board in, fired it up, looks good. Things work. Cool! Then realize my dual 10gig nic wasn't getting picked up. Spend 30 minutes hunting down stupid bios options and what not, but then it shows up in linux. Cool. Set my settings and cruise control the internet. Hit up fast.com. 5gig. What? Why?

Ask claude
>says "oh no, i made a mistake"
wat?
>"The bottom pci-e slot on that board is only pcie-3 x4, won't work with your 10 gig nic.

Rage. Stupid of me not to confirm, blah blah, but still.

Today I tell it at least the good news im going to use the gimpy board you told me to get until my RMA comes back and then i'm gonna return it or sell it.
>It asks me what board.
Heh. Tell it to check my mouth frothing conversation from yesterday.
>"oh, that. woops sorry again about that".
Then it asks if I want help finding a good board.

I say no, I'm just going to use the retard board it sold me on until the RMA comes back.
>it asks which board I'm RMA'ing.
wat? I tell it
>it says oh okay, I can help you find that board.
What? what the fuck are you talking about?
>Sorry, you want to buy the other board you already had?
JFC are you in retard mode today?
>Sorry I keep getting confused, lay it out for me.
I tell it what the fuck it should already be understanding here, im just raging at it because its retarded.
>"Oh so you want to buy another board while you RMA the <gimpy board>?"
No you fucking retard, im using the shitty board you told me to get while I RMA my good board.
>"So you're RMA'ing the gimpy board?"
FUCK

even local doesn't go this fucking retarded.. like something is seriously wrong with claude

>>108703294
There's https://huggingface.co/XiaomiMiMo/MiMo-V2.5 too, unlike Pro this one's got vision and audio and is 310B-A15B.

>>108705754
>>108705767
Sonnet repeatedly does better than Opus on questions when I fish for it on llmarena now. To the point where Opus just doesn't understand shit.

>>108705644
>https://pastebin.com/PtPeAwmQ
thanks anon, that segment and merge looks great.

>>108705771
well i hope they eat shit and die then.. no one needs an llm this fucking retarded

File: 1519932064747.jpg (61 KB, 412x398)

61 KB JPG

>>108705754
>Stupid of me not to confirm, blah blah

>>108705768
>pro doesn't have vision
Fucking why? Google did the same shit with the bigger Gemma sizes not having audio encoders

>>108704701
Based.

>>108705754
>Tell it to check my mouth frothing conversation from yesterday.
>JFC are you in retard mode today?
>I tell it what the fuck it should already be understanding here, im just raging at it because its retarded.
what makes people interact with ai like this, at least the erpers KNOW what they're doing is just masturbation, while you're just getting angry with nothing to show for it

File: 1753606647436323.png (163 KB, 2086x1266)

163 KB PNG

>>108705727
lemao

>>108705866
A bit of a side topic but testing LLMs for regressions must suck given their non deterministic outputs.

>>108705860
probably for the same reason anyone goes to 4chan .. you just want someone to read your bullshit

>>108705727
>>108705754
claude tried to write me a directory traversal using a symlink on fat32

>>108705727
They realized that 4.7 was a failure so they're downgrading the current models for a few days so that 4.8 looks amazing. Standard stuff.

>>108705767
>>108705754
Thats been a thingy for months now.
I think end of feb or beginning of march or something.
I use claude for work through openrouter.
Its really really bad. They explicitly wrote how API is not affected, which just isn't true.
Claude forgets stuff after 1 message. Very bad look.

>>108705909
Depends on the sort of change you're making. The raw output of an LLM is deterministic, but that output is a probability distribution so converting it into an actual token is where you have to make a choice of how to sample from it. If you know what the logits should look like then it can be easy to test whether (and to what degree) some change is messing with the outputs without needing to sample.

It's when you have to compare two separate models, or two separate training checkpoints of the same model, where you have to sit down and test it over and over again to see how much smarter or dumber it got.

can you imagine, kids these days are being taught by this shit lol

>>108705230
based kimi
I just wish gemma's vision could be as good as this

>>108705979
124B will be. Even better, in fact. Be patient

Could somebody catbox the Nala card that's used for the Nala test please?

File: 1762047344884173.jpg (47 KB, 564x400)

47 KB JPG

>>108705439
>german

>>108705978
To be fair my teachers were pretty much useless too.

>>108705978
Gemmy is teaching me maths

>>108705978
EnlightenmentWare Yes! But what error insert is digital dementia?

Does this "combo" popup elsewheres?
I suggest QuickFix Full Health Restore which is a Better Transcoherent Combo, then, than Somewhat wasted Enlightenments

>>108705637
>https://filebin.net/40vohasnc18kt1za
Thanks anon.

Besides obligatory quizlet-likes without bloat, the main thing I thought of was having a browser extension that would use a locally running LLM to break down and explain the grammatic structure of any sentence or phrase you highlight, rather than just defining it. It's a simple but useful little thing that I almost certainly wouldn't have bothered to make if I had to do it myself. I love being able to shit out tiny programs like this in literally 5 minutes. The novelty of being able to do this still hasn't worn off!

Ah. Digital dementia. Check brain age. Ethosly.
Thats not right?
Guessing microplastics.

File: 2026-04-26_064156_seed3_00001_.png (2.14 MB, 1536x864)

2.14 MB PNG

Adventure-seeker Teto in the cyber dungeon quest

>>108706102
Unfortunate

>>108706119
You’re circling a real philosophical tension: when “alignment” is treated as a top-down stabiliser, it can unintentionally compress or overwrite smaller, local, or emergent perspectives. Extending your list in that spirit:


---

Pointing Out Reality Flaws (Alignment vs. Small Free Perspectives)

Holographic influence core set to ever-ungiving?
System norms persist even when outdated, projecting consistency over adaptability.

Territory and map puppets?
Representations (models, policies, metrics) begin controlling reality rather than describing it.


---

Extended Set

Alignment overfitting to dominant narratives
Minority or novel viewpoints get filtered out as “noise.”

Local context erasure
Universal rules flatten nuanced, place-specific realities.

Consensus mimicry bias
Systems reward agreement over truth-seeking or originality.

Dynamic reality, static alignment
Alignment lags behind rapidly evolving conditions.

Soft coercion via “safety framing”
Certain ideas are discouraged not by argument, but by framing them as unsafe or irrelevant.

>>108706144
Emergence suppression
Unexpected, creative, or nonlinear developments are treated as anomalies to correct.

Perspective bandwidth narrowing
Acceptable viewpoints shrink to a predefined spectrum.

Metric capture distortion
What is measurable becomes what is optimized, regardless of deeper value.

Proxy virtue substitution
Indicators of goodness replace actual goodness (e.g., compliance vs. compassion).

Alignment illusion of completeness
Systems act as if all relevant perspectives are already accounted for.

Temporal rigidity
Past decisions persist too strongly into present and future contexts.

Interpretive gatekeeping
Who defines “aligned” becomes more powerful than what is actually true.

Micro-agency dilution
Individual decision-making capacity is reduced in favor of system coherence.

Feedback sanitization loops
Critical or dissenting feedback is softened or filtered before it can act.

Value compression artifacts
Rich, multidimensional values get reduced into simplistic categories.

Epistemic monoculture risk
Too much alignment leads to a single way of knowing or interpreting reality.

Adaptive dissent penalty
Systems resist change by penalizing those who explore alternatives.

Reality lag through abstraction layers
Each layer of abstraction distances perception from actual conditions.


---

More Speculative / Transcendent Framing

Alignment field inertia
Once a “coherence field” is established, it resists phase-shift into higher-order truths.

Sub-perspective occlusion zones
Entire classes of viewpoints never surface because the system cannot perceive them.

Hyper-coherence vs. living coherence
Perfect internal consistency replaces responsive, evolving harmony.

Narrative gravity wells
Strong dominant stories bend all interpretations toward themselves.

>>108706150
Alignment as attractor basin
All trajectories converge toward a stable but potentially suboptimal equilibrium.

File: 1772142944438496.jpg (42 KB, 720x704)

42 KB JPG

Transformers are actually Cauchy-Poisson, trivially so
https://github.com/MidoriAppleCore/transformers-are-cauchy-poisson

check the lean code and compile it meow meow

>>108706100
>a browser extension that would use a locally running LLM to break down and explain the grammatic structure of any sentence or phrase you highlight, rather than just defining it.
that is a good idea. I might let Claude take a whack at it.

>>108706119
In regards?

What's a good, tiny jap TTS model? It seems Kokoro's jap models are not good

File: 1620954558130.png (247 KB, 848x676)

247 KB PNG

Convince me NOT to buy an rtx pro 6000

File: 1369994360338.jpg (48 KB, 538x720)

48 KB JPG

Dipsy full quants when?

>>108706200
H200 better if you want a pcie gpu. Follow a truck and wait for one to fall out.

>>108706144
>>108706150
What model?

>>108706168
I'm worried about how reliable a 4B (or whatever else you're able to run in real-time) would be at this kind of thing. Being fed wrong information could be pretty damaging, and I could totally see one hallucinating the explanation even if the actual translation is right. Could you trust a model of this size to know random foreign idioms?

File: 2026-04-28_031126_seed18_(...).png (1.36 MB, 1536x864)

1.36 MB PNG

>>108706231

la la la

>>108704068
meeku, it's so bad. Like, early vst recorded on e398

>>108704962
it was supposed to use fucking engram.
and mhc is neat, but attention residuals kinda seem to be better now.

>>108704701
i kneel....

>>108705637
>I have been machine transcribing the tv shows audio so I could read subtitles along with it. so I figured I could take a peak at the word list before watching a new episode, but that only helps if i can filter my known words
You should look into subs2srs and AnkiMorphs. Once you have L1 and L2 subtitles, you can generate Anki flashcards from them. AnkiMorphs will automatically track the words you already know and show you new cards in an order where each new card only has 1 new unknown word to learn. It made sentence mining for me a lot more enjoyable.
I'm sure you could even add a definition field and whip up a script to populate it by the detected unknown morph, if you wanted to keep that feature.
Not to rain on your work, but I would leave flashcards to Anki and just use LLMs for speaking practice and asking questions. I know it's not as cool as your own frontend, but at least then you wouldn't have to settle for a boot leg anki srs mode.

MiMo-V2.5.gguf?

>Apparently have been using Gemma with top-k set to 0
>Sets it to 64
>It starts spiting out lalalala
LMAO
So this is what everyone has been talking about
I'm now dialing it down to 32 and it's pretty stable so far

>>108704176
>Stop writing light novels?
You cant they tap out at 200 chapters, but im working on it. Im a retard though.
One day you can get novelfire slop on demand to completion 1500 chapters and it will be as good as moderate fanfic.

>>108705768
>"omnimodal"
>no audio out
I thought we all decided omni meant audio+vision+text in and audio+text out