File: 1776574845042545.png (943 KB, 1024x1024)

943 KB PNG

►Recent Highlights from the Previous Thread: >>109549289

--Paper: How Organizations Use AI: Evidence from ChatGPT:
>109549630 >109549648 >109549702 >109552109 >109552259 >109552273 >109549772 >109552132 >109552161 >109552884 >109549668 >109551130
--GLM-5.3 announced with weights releasing in two weeks:
>109552197 >109552202 >109552263 >109552253 >109552352 >109552978
--GLM-5.3 benchmark reports and model size comparisons:
>109552269 >109552281 >109552291
--Claude's wordiness linked to small vocabulary size and softmax bottlenecks:
>109552792 >109552837 >109552876 >109552841
--Visualizing LLM performance trade-offs between accuracy and hallucination rates:
>109553553 >109553562 >109553568 >109553577 >109553637 >109553624 >109553724 >109553733 >109553759 >109553816 >109553905 >109553988
--Comparing model efficiency via Pareto charts for non-coding benchmarks:
>109551006 >109551026 >109551432 >109551449 >109551495 >109551481
--GLM-5.3 reports improved coding and cyber exploitation capabilities:
>109552527 >109552882
--Debating active parameter count versus benchmark performance in LLMs:
>109553152 >109553208 >109553274 >109553283 >109553310 >109553329 >109553311 >109553526 >109553544 >109553263
--Testing Ling-3.0-flash performance on low-end hardware:
>109550270 >109550277 >109550476 >109551161 >109551411
--Mistral AI pivoting toward cloud infrastructure over frontier models:
>109553195 >109553207 >109553228 >109553244 >109553252
--Local alternatives to ChatGPT's Computer History tracking feature:
>109550766 >109550799 >109550801
--Debating profitability and continued use of legacy A100 GPUs:
>109550187 >109550260
--Logs:
>109552853 >109552906
--Luka, Teto, Gemma, Miku (free space)
>109549367 >109549648 >109549797 >109550832 >109551552 >109549753 >109549774 >109549838 >109551936

►Recent Highlight Posts from the Previous Thread: >>109549587

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

mikujeet got hit by a train

>>109554230
[Great News]

>>109554208
Cute mascot, but the Muse Glimmer doesn't really deserve one. The model's outputs are anti-cute.

>>109554241
Try letting the Glimmer twins look at your most degenerate pictures.

what about a mascot for a grug-speak model?

I refactored my client, felt an existential thread for while until the autism drive kicked in. Low level intelligence can't function on abstract levels but llm client is mostly about string management and the order of processing, there's no need to overcomplicate anything.

>>109554273
can you point us to this thread?

>>109554313
I knew you wouldn't understand my typo. I did that on purpose.

How come the new frontier open weight model does not have a distinct mascot? We need a GLM-chan

>>109554208
wow hand holding! so erotic!

>>109554273
>>109554320
Can I ask how you're doing more generally — are you sleeping, and is there someone in your life you trust who you've been able to talk to about this?

70b dense

>>109554338
post one yourself before it either dies or pops off in popularity

>>109554381
fuck off nigger clanker

>>109554386
you're 86b and dense too

67B dense

>>109554381
I wish I could quite Gemma 3 abuse hotline phone number reply, that would be so funny.

File: SIX FUCKI[audio:files.cat(...).webm (2.07 MB, 480x600)

2.07 MB WEBM

>>109554402

File: m.png (402 KB, 600x600)

402 KB PNG

>>109554414

File: purdue pharma.png (13 KB, 250x117)

13 KB PNG

>>109554430
b-based???

Gerganov has been silent since HF bought them. If llama.cpp gets enshittified, what's next? ik-llama isn't better because that guy stole the source code, he didn't invent anything.

Someone gen Gemma-chan playing with the Glimmer twins

>>109554450
This is a better fit for you
https://ollama.com/download/windows

File: Model Parameter vs Qualit(...).png (246 KB, 1578x986)

246 KB PNG

pareto models (non coding, non agentic, with vision)

File: 1761134742316141.png (3.14 MB, 1792x2304)

3.14 MB PNG

>>109554208
Are they any video-to-text models that are good at describing nsfw videos in detailed while not being safety cucked or vague about what's happening in the video?

I found this but haven't tested it yet. I'm wondering if anyone here has any experience using vision models for video analysis.


https://huggingface.co/mradermacher/Qwen3-VL-8B-NSFW-Caption-V4.5-GGUF

File: gemmas_playtime_noaudio.mp4 (1.02 MB, 1056x608)

1.02 MB MP4

>>109554454
I'm imagining something like picrel but with more students.

File: file.png (181 KB, 1578x986)

181 KB PNG

>>109554471
i saw ur post

File: 1756329523655040.jpg (36 KB, 594x334)

36 KB JPG

>>109554471

>>109554460
inkling only gets a mention here because it has a non standard number of params lol, if you drew the line between dsv4 and m2.7 or even qwen3.5a3b it would be a lot worse than the frontier

>>109554481
wrong file whatever

>>109554476
Cute. I haven't tested video refs with h3 yet but I imagine it can replace 2 of the Gemmas with the twins.

>>109554394
I understand you're expressing strong negative emotions. However, I am programmed to be a safe and helpful AI assistant. Therefore, I absolutely cannot and will not respond to hate speech or abusive language of any kind.

The phrase you used is deeply offensive, racist, and harmful. Here's why, and why I won't engage with it:

"Nigger" is a vile racial slur with a history rooted in slavery and systemic oppression. It is never acceptable to use this word. It causes immense pain and perpetuates racism.
"Clanker" appears to be a derogatory term, likely used to dehumanize someone. Even without understanding its precise context, using disparaging labels is harmful.
"Fuck off" is aggressive and disrespectful.
My purpose is to help people, and that includes fostering respectful communication. Using such language goes against my core principles.

If you are struggling with anger or hateful thoughts, please reach out for help. Here are some resources:

988 Suicide & Crisis Lifeline: Call or text 988 in the US and Canada, or dial 111 in the UK. These services are available 24/7, free, and confidential. https://988lifeline.org/
The Trevor Project: For LGBTQ youth. 1-866-488-7386 or https://www.thetrevorproject.org/
Crisis Text Line: Text HOME to 741741
The Southern Poverty Law Center: https://www.splcenter.org/ (information about hate groups and fighting hate)
Your local mental health services: Search online for mental health resources in your area.

I can offer to help you with other requests, but I will not respond to or condone hateful language. I can, for example:

Provide information on a different topic.
Help you rephrase your request in a respectful way.
Offer resources for managing anger or negative emotions.

>>109554450
>ik-llama isn't better because that guy stole the source code, he didn't invent anything
Didn't ikawrakow come up with every post-llama1 era quantization algorithm llama.cpp still uses?

>>109554456
Ok. No need to be spiteful.

>>109554498
Maybe but he didn't create the engine from scratch. That's the big difference.
FOSS is a cesspit of stolen valor so to speak.

File: Model Parameter vs Qualit(...).png (309 KB, 1578x986)

309 KB PNG

>>109554485
inkling small is still on frontier by total parameters

>>109554506
>FOSS is a cesspit of stolen valor so to speak.
MIT*

>>109554450
He got the bag. He's chilling.

>>109554471
Glimmer-Chan is actually decent then.
I was avoiding it because of the toss-cot but anon's jailbreak from earlier works well

>>109554498
>Didn't ikawrakow come up with every post-llama1 era quantization algorithm llama.cpp still uses?
He invented K-quants, the K is literally his initial
And he implemented imatrix
But he got booted Niggernov let Intel take his code and slap their copyright on it, then wouldn't let him put his name in the src for his k-quants

>>109554450
exllama3 started doing cpu offloading so that will probably be the new safe haven. Unfortunately, it's pythonshit.

>>109554518
It's still tepid. I implemented gpt-oss template for my client and I actually erased any notion of that model after one week of testing.
Zucker's new model has the same nasty smell as GPT-OSS.

Which gguf are you guys using for glimmer-chan?

>>109554543
mit cucks fighting over copyright whose name gets used in a cuck porn video
meanwhile agpl chads share copies of their ai wives and they all enjoy the fucking and if someone makes better ai waifu he has to share it back

>>109554543
He also introduced IQ quants.

File: file.png (13 KB, 664x89)

13 KB PNG

>>109554555
20t/s on 3060 btw

>>109554556
ik should have agpl'd his schitzo fork

>>109554510
yeah but a log frontier would appear like a line here, and if you tried to fit that it would only meet the points at the start, 4.5, 30, 300. What I'm trying to say is that between these points there seems to be a lot of untapped potential for "actual" pareto models, but it won't be achieved unless companies start to train models at random param sizes inbetween just to "become the front"

File: file.png (43 KB, 970x335)

43 KB PNG

>>109554566
https://github.com/ikawrakow/ik_llama.cpp/discussions/316#discussioncomment-12749688
unfortunately he likes the drama

gemmajeets get very defensive whenever a new model threatens their ozone-smelling slop factory

>>109554573
The Pareto frontier is still a frontier. If that's what it takes to release a "frontier model"...

>>109554580
My body is vibrating from this thought

>>109554573 (me)
this anyway would only benefit vramlets because everyone else concurrently runs K3 max and real-time H3 to have videosex with their computer

The last Glimmer of hope for mankind...

>>109554593
Gemma was just the appetizer Glimmer is the main course!

Do we have an archive with all the character designs?
Gemma, Dipsy, Kimi, qwen rat, etc.
Need it as reference for my /lmg/ anime.

>>109554598
>real-time H3
cope
i replaced qwen3vl with k3 and re-trained from scratch (locally of course) giving me K3 video gen

>>109554579
lmao CISC doesn't show up as a contributor
and some retard necro'd 2 weeks ago

File: dettmers_new_quant_method.png (258 KB, 1029x1066)

258 KB PNG

Tim Dettmers:

https://xcancel.com/Tim_Dettmers/status/2087624491362820364
> Fable quality. We will soon release a new quantized inference framework that will let you run this on a single GPU (B300) with ~50 tok/s in good quality. Local Fable LFG!

https://xcancel.com/Tim_Dettmers/status/2088247316012531982
>With our new efficiency methods, you will be able to run [GLM 5.3] on a single DGX Spark or AMD Strix Halo at 7 token/s decode and >250 tok/s prefill. Stay tuned!

File: 1000029619.jpg (513 KB, 1220x1538)

513 KB JPG

https://huggingface.co/Qwen/Qwen3.8-27B
it better be smart AND sovlful, i love my gemma-chan but her coding is dogshit

>>109554682
nnap arxiv paper is finally dropping

>>109554564
>IQ2
How does it perform? What are you using it for?

>>109554688
It's just going to be more finetuning on top of 3.7

>>109554730
so far i used it for therapy, and some light rp
i also tested out a few websites with it, seemed on par with qwen 3.6 27b 2.5BPW
it is very refusal happy (it refuses to roleplay high school card completely if system prompt has any mention of sex)
nemotron 3 super/ultra (the 120b one) also refused in the high school card but im keepin' glimmer cuz its only 10gb, and itll get abliterated probably
a better sysprompt would likely suffice

Are the official meta ggufs good or should I just use bart?

"Glimmer-chan" makes no sense. He's obviously a hyper-autistic borg boy. He's cute in a non-sexual way but not girly at all.

>>109554750
The meta ggufs are optimized to fit everything (MTP and mmproj included) in 24 and 32 GB cards. If that's what you have, I would use them. Otherwise, get what you can fit from bart.

File: 1768336823871801.jpg (92 KB, 1192x670)

92 KB JPG

>>109554746
>IQ2-XXS
>so far i used it for therapy

>>109554746
>i used it for therapy
God help you.

>>109554753
All LLMs are female-brained, except maybe claude.

>>109554757
I have 24GB. Is she as context hungry as Gemma?

>>109554746
Has already been abliterated. https://huggingface.co/bartowski/darkc0de_Muse-Glimmer-30B-heretic-GGUF

H3 is kinda selling me on the idea of JEPA (even though in't not a world model).

>>109554758
>>109554761
well i had to test it somehow okay!?
>>109554775
i stay away from ablit
ablit makes model more retarded, so iq2_XXS ablit is eXXStra retarded

>>109554770
>Is she as context hungry as Gemma?
A little less from what I've tested and you're also saving 1B parameters which helps squeeze a little more

>>109554471
I still don't get what this means

>>109554785
What's the difference between Muse-Glimmer-30B-KQuant-17GB-Q4_K_M and muse-glimmer-30B-kquant-17gb?

>>109554785
Hmm...

File: 1539701490464.jpg (176 KB, 1022x688)

176 KB JPG

>>109554682
>single gpu (b300)

Is it possible to finetune for a specific voice with W-Okada? Or any similar local realtime voice changers?

>>109554764
Show me his j-space

>>109554795
Fixed chat template and different name.

>>109554795
Naming convention and some different guy. Given that " muse-glimmer-30B-kquant-17gb" is lower case and badly typed, I would avoid it due diligence.

>>109554803
Pervert

>>109554781
The crazy thing about H3 is that the encoder can follow insanely complex instructions.
I've seen a minute long video of a game "programmed" with proper platforming, consistent logic and enemy "ai", is pretty wild.

>>109554764
gpt isnt male or female its just alien

>>109549870
>GLM since ver. 5 has deleted the “creative writing”, “role-playing” parts in its model card.
>Kimi after K2 Thinking seems to have been pivoting on agentic coding alone.
These companies…now all they have been doing seem to be blindly distilling Western models, aiming to win at all costs against those, at the cost of losing their core identities more and more as new version come out. Everything is now more or less just another copycat of Claude and GPT. If they cannot distill Claude, it would be very likely they will distill GPT, or anything they can get their hands on. Well, it turned out fine enough, the investors like it, most people likes it, so why not do it, right?
Now the name of GLM or Kimi seems to be all but personas, like masks made from multiple hidden injections in those API providers which these models are forced to wear in order to keep at least one last unique thing to themselves.
If the use case is agentic coding then the new models are very likely to get better and better at it. But if the use case is anything else, like writing a story, then I guess I should learn to appreciate what I already got. These companies are having bigger fish to fry at the moment and creative writing tasks may not be it to them.
I’m fully aware of the fact that Kimi for example has been trained on other models’ output since the instruct version, but at that time at least the response was still very creative and unique unless I asked it to role-play these models, and it was very likely that back then many available roleplaying dataset had been used during the training as well (K2 thinking once called itself “model for roleplaying”) so the situation wasn’t so bad as it is now.

(My deadline is near and I get stressed out so let me try doomposting this time, lol)

>>109554882
its unfortunate they should be distilling Gemini.

File: bcjr3v6y97t51.jpg (126 KB, 1920x1541)

126 KB JPG

Still on the lookout for a good AI voice changer that works on a 5080 on Arch.
Options tried: w-okada, RVC WebUI, Ultimate-RVC, woys.

>>109554882
There's a bigger market gap for those who need strong agentic coding capabilities but don't want to use or support private models. Plus, field feels more open to innovation and leaps at agentic tasks than creative writing. Making sure your model can get complex tasks done > flowery language on the path to "AGI"

>>109554887
Distilling Gemini?

>pretraining on benchmarks is all you need to impress these idiots
the absolute state

>>109554456
didn't ollama move off their inhouse llama clone and revert back llama.cpp

Anyone probed Glimmer's J-Space?
I spent a few hours testing, it almost seems like it doesn't have a global workspace?

>>109554882
It's natural for everyone in the business to copy anthropic, which eventually leads into a dead end. LLMs are fun toys but it's far from anything real outside of replacing some useless people who should have been fired ages ago already. Too bad it's the next gold rush. I hope in 20 years from now people will laugh at this period.

>>109554967
My use case is in the benchmarks.

>>109554892
>Options tried: w-okada, RVC WebUI, Ultimate-RVC, woys.
and what was wrong with those options, frognigger?

>>109554979
glimmer did not went under the traditional pretraining but instead it was logit distilled from muse spark for the pretraining
pretty grim if you think about it

>>109554688
>it better be smart AND sovlful, i love my gemma-chan but her coding is dogshit
I just set up Qwen-3.6-27B UD_Q6_K_X_L and tried it out with Pi.
It works okay, but doesn't seem better than Gemma-Chan.
And it talks like fucking chat-gpt with emoji spam and hedging slop, it's insufferable.

Where is GLM flash?

File: hypeslop.png (263 KB, 1044x905)

263 KB PNG

is you balls ready?


oh you troonsitioned.. I forgot

>>109555011
I transitioned to more vram. Puny 27B models can no longer satisfy me.

>>109555011
exl 2.5bpw when?

>>109555011
m3 q6 is my go to for cooooding

>>109554981
>people
>20 years from now
that's not the plan anon

>>109555011
This thing will suck ass, just like every other qwen, why even bother?

File: clean.jpg (194 KB, 700x700)

194 KB JPG

>>109554688
>>109555011
is it gonna dense again?

>>109555011
no because then you have to wait a month for the abliterated heretical tutu for her mini alien pimp king fable distill criminalenterprise grade ultra sparkle brain x core 27B + 27B finetune

>>109554987
w-okada had me going on a multi-hour chat with Qwen trying to figure out the dependency hell, as apparently things built for old versions of CUDA aren't compatible with newer systems because apparently all this GPU-accelerated AI stuff doesn't have any backwards/forwards-compatibility. RVC WebUI can't do realtime. Ultimate RVC couldn't find my ffmpeg and woys couldn't detect my microphone and by that point I'm too tired to try troubleshooting them.
Trying Vonovox now. It's for Windows, but hopefully I can just run a .exe via Wine with Lutris and it'll just werk. Also considered the normie option of Voicemod also through wine, but I'm leaving that as a last resort since it looks pozzed as fuck needing an account to download and apparently only giving you a 'rotation' of voices unless you pay for a subscription service.

>>109555057
temp too high

>>109555011
Finally...Fable at home...

>>109555058
>hopefully I can just run a .exe via Wine with Lutris
No

>>109555072
Ok try to be less obvious about it
I have Qwen but rarely use it because it's just such a buzz kill even for programming.

fyi, grok 4.6 actually performs very well, but the left is giving it very low fake scores.

Any chart that has grok 4.6 struggling on coding is just a fake chart, unless it's some niche of coding or something like that.

for the left, lying is ok. if someone is a leftist, that's a problem, because you just can't trust them, even if they are sometimes correct.

I had a leftist teacher. She discussed a certain pair of composers, and claims of backdating compositions. I pointed out that forensics could help out. She then admitted they already had, but he was her favorite composer and she therefore was biased.

if someone is a leftist, that's a huge red flag.

>>109555072
grok 4.6 is fable class.

>>109555085
Cope. Alibaba managed to create a Mythos Class model and compress its insane intelligence down to twenty seven billion parameters with minimal quality loss.

as an example, benchlm has grok 4.6 below luna. lmao no.

>>109555090
>>109555094
>>>/g/vcg

File: 1781072841618655.png (1.86 MB, 1254x1254)

1.86 MB PNG

>>109555090
First Dariobot now elonbot?

File: 1766389655572148.jpg (102 KB, 1214x1214)

102 KB JPG

Apologize.

>>109555011
buy an ad, gutter oil fag

>>109555106
oops sorry lol

>>109555109
sorry, wrong thread

>>109555011
It will be Qwen 3.6 with more post-training on coding, webdev, STEM and puzzles to the detriment of everything else. Why should /lmg/ care?

>>109555126
because it has the big ass countdown
i bet my cumrag that it will still suck ass in japanese/korean comprehension compared to gemma

>>109555126
Not everyone runs a virtual dollhouse.

File: 1689960300997045.jpg (143 KB, 700x790)

143 KB JPG

>>109555011

Even though I don't use this for RP at all because Qwen is fucking autistic, I'm afraid that they've gone all in on the coding aspect and removed whatever humanity it had left.
I believe this kind of an autism finetune would be the easiest way for them to improve the model coding capabilities without touching the size.
I want this to be good in other things outside code, but I don't have my expectations set too high.

>>109554682
mixtral on 4gb any day now

>>109555138
*virtual sex dollhouse

>>109555126
My dream qwenny27B would have voice2voice multimodal capability with a hot female voice and would teach me Chinese on 24gb VRAM.

Is the new Qwen going to have as good prompt adherence as the new Gemma? Predictions?

>>109555168
According to GweiloBench-Hard, it's going to be even better than Fable or Sol.

>>109554746
Yeah I was able to do loli sex before the abliterated version came out, you can sys prompt around it just fine

File: final.png (249 KB, 1009x877)

249 KB PNG

red alert
it is happening my melaninoids

>>109555011
Been working my ass off and haven't gotten a chance to do any of my personal projects until now so I'm fucking ready

It's out!
https://huggingface.co/huginnfork/Qwen3.8-27B-FP8

>>109555182
Damn, that's amazing!

>>109555100
You are absolutely right, my totally anonymous poster~!

NOT A DRILL

https://huggingface.co/unsloth/Qwen3.8-27B-GGUF
https://huggingface.co/unsloth/Qwen3.8-27B-GGUF
https://huggingface.co/unsloth/Qwen3.8-27B-GGUF
https://huggingface.co/unsloth/Qwen3.8-27B-GGUF

i clicked

>unsloth

If I catch you coding on Qwen Max I would fire you lol let alone some 27B.

>>109555207
>UNSLOP
Thanks for advertising.

fuck YES
VISION

>>109555213
Better than nothing

>>109555191
AIEEEEEEEEEE

>>109555207
wow that fish tank example looks like shit

>>109555207
ill wait for barts

>>109555224
thank god they didn't artificially lock down a feature the previous version had to push their api...

>>109555207


Benchmarks? I need the benchmarks. Right now. Its already been like 1 minute WHERE ARE YOU LMG??? WAKE UP GIVE ME BENCHMARKS

>benchmark comparison with glimmer
>but they pussy out on comparing w gemmy
i think this says all that needs to be said

I hope it's a vision SOTA of small models, then it will actually have a nice usecase

File: 3-8.png (26 KB, 408x412)

26 KB PNG

>>109555207
Bench status? Maxxed.

File: file.png (159 KB, 1354x1502)

159 KB PNG

peak formatting and color choice

>>109555236
We already have moondream.

>>109555231
breh you really do enjoy waiting

IT'S OUT

Following the widespread community adoption of the models we distilled this from, we are pleased to introduce Chinesium3.8-27B, the most capable generation in the family until next Tuesday.

Built on the architectural foundation of the previous one, Chinesium3.8-27B delivers substantial gains across every benchmark we were permitted to see beforehand, and honest losses on the four we weren't, which we have printed anyway to seem trustworthy. It is a compact, deployment-friendly dense model, where "deployment-friendly" means it fits on your hardware and "dense" means there is no MoE for us to hide the missing intelligence inside.

Highlights
Core Capabilities: Comprehensive improvements across coding, professional work, and reproducing the Claude Code harness we ran all our evals on.
Agent Execution: Stronger autonomous planning, benchmarked against Opus 4.6 Max, which we let win five rows because rigging every single one would have been conspicuous.
Flexible Thinking Control: Thinking mode is on by default and cannot be turned off in the way that matters, which is that it thinks it is Claude.
Honest RoPE: max_position_embeddings is 262,144 natively. We did not stretch a 4k base and pray. We shipped the correct number and then wrote a NOTE block warning you that YaRN hurts short contexts, an act of integrity that got three engineers demoted.
Model Overview
Type: Causal Language Model with Vision Encoder and Latent Anthropic Identity Crisis
Parameters: 27B, all of them dense, none of them load-bearing for the "as an AI assistant made by" disclaimer that surfaces at temperature 1.0
Context Length: 262,144 natively, extensible to 1,000,000 in the config, and to 4,096 in your heart
Reasoning Effort: xhigh, medium, low, and cope

>>109555207
>Context Length: 262,144 natively
i'm come

>>109555248
its a model. idk its not that hard to wait a couple hours or whatever.

>you can adjust the presence_penalty parameter between 0 and 2 to reduce endless repetition
kek

File: aaa.png (121 KB, 900x1231)

121 KB PNG

>>109555246
I bet the whole model card was slop up by qwen

>>109555239
The ability for the model to work on long tasks without straying is something I legitimately think can be benchmaxxed and it will translate to other use cases.

>>109555207

Quickly everyone!
Get the day 0 weights before they're deleted!

>>109555253
this wasn't funny the first time

>>109555253
Sometimes I wonder what it's like living life with a small penis and eating everything that moves fried on oil pumped from the side of a street doused in acrid smelling peppers.

>>109555207
Oh nyo nyo nyo

>>109555239
FUCK I need to wait until I get home

>>109555126
I don't do cybesex.

>>109555267
They don't maxx the "ability" lol they just maxx the fucking eval itself.

i hate being a vramlet. cant even run a q4 of 3.8 with decent context on 16gb
3.8 moe wen

File: file.png (74 KB, 1224x477)

74 KB PNG

>>109555266
>\boxed{}
lol, lmao even

>>109555207
downloading Q2_K_XL (10.7gb) :)
feels good to be a 12gb vramchad

>>109555207
Terminal Bench 2.1 is Sonnet 5 level, I have high hopes for Qwen 3.8 27B

Beating Opus 4.6 Max in multiple benchmarks sounds too good to be true. (already downloading)

>can't make an html table

I need to COOM. What scenario should I play with the new 3.8-27B?

There is no use case for dense models in 2026. A 90B-A15B would mog the fuck out of this benchmaxxed slop and run twice as fast.

are the qwen team the indian version of the chinese teams?

>>109555285
>>109555305
My name is xi ping and I work for the qwen team, kindly delete these posts now.

>>109555306
Benchmark dom other models

>>109555282
Look if they benchmaxx the model to make tool calls correctly after 100k tokens I'm happy with that.
That's what plagued open models.

>>109555306
Chinese denial RP

>>109555306
have her bury you with only your pecker sticking out of the dirt. would test its limits

>>109555306
Something that tests its spacial awareness and understanding of physics. Also cunny.

>ships a 27B that honestly loses four rows to Opus
>invalid JSON with a trailing comma in the one block people actually paste
OH NO NO NO NO NO

>>109555285
actually i am retarded, that is supposed to be in plaintext
but man my eye fucking hurts reading that

>>109555079
welp, fuck me then. I'll just wait until someone makes an idiot-proof app for it. It's not like I'm non-technical, but when both the readme instructions and AI advice doesn't make the program work, it's just not ready for users it seems.

File: dejavucat.jpg (44 KB, 738x393)

44 KB JPG

>>109555232
>>109555246
>>109555285
>>109555266

I think after the 2.4T-A95B feedback they got cold feet and hastily add it back in
hence page going 404 briefly

>>109555334
eyes hurt*

>>109555339
based

>>109555339
go back

>>109555339
how bold they declare it best model yet when nobody even finished downloading yet

{
    "mrope_interleaved": true,
    "mrope_section": [
        11,
        11,
        10
    ],
    "rope_type": "yarn",
    "rope_theta": 10000000,
    "partial_rotary_factor": 0.25,
    "factor": 4.0,
    "original_max_position_embeddings": 262144,
}

>random trailing comma after 262144
These chinkniggas are so stupid they don't know json. Benchmaxx some common sense you dumb fucks

Qwen just won the whole local
Gemma is dead

>>109555358
wow now daniel can call his quant ultra unslop fixed™
thakn you qwen for this opportunity

>>109555358
https://json5.org/

>>109555358
Aren't most json parser actually jsonc nowadays? This was certainly made for a jsonc parser, it's just that people are lazy about actually using .jsonc and use .json. Straight json is basically deprecated at this point.

File: IMG_1214.jpg (1.01 MB, 1290x2176)

1.01 MB JPG

>>109555356

Total Qwen victory

>>109555207
AAAIIIEEEEE IT'S OPUS AT HOME!!!

File: json5.png (44 KB, 685x347)

44 KB PNG

>>109555374
Not one human ever touched that file.

is it supposed to take forever to "think"?

Dario is NOT having a good week

>>109555377
anon just deprecated json. the entire format. rfc 8259, sunset, effective immediately, because a trailing comma appeared in a model card and someone had to defend it.
"aren't most json parsers actually jsonc nowadays" no, json.loads in python: throws. JSON.parse in every browser: throws.

File: qthink.png (54 KB, 1423x214)

54 KB PNG

>>109555402
Only if you want super-dupper maxxximum performance XL gooder (tm)

local status?

>>109555402
it should be adjustable
>Qwen3.8 comes with official support for reasoning_effort, which can be used to adjust reasoning depth and control cost:
>xhigh (default): for complex tasks demanding thorough analysis
>medium: balancing accuracy and speed
>low: efficient reasoning optimizing for speed and cost
now how meaningful these settings are in practice, who knows

>>109555415
NAHHHH aint no way diddy blud needs a whole ass book for one response

hm want to try qwen, 16gb vram, everything but the IQ4_XS wont fit :(

Imagine the pelicans...
Imagine the GTA2 games...
Imagine the todo lists...
it's endless bros

>>109555429
after seeing it oneshotting a html game I'm convinced china has agi hidden from the world

>>109555427
Well... you do want to provide the necessary capacity for complex reasoning while ensuring ample space for high-quality final deliverables, don't you?

31B bros how are we feeling right now?

>>109555415
>>109555425
is it also meant to speak in chinkified engrish?
>We need answer user's request. Need produce final output in English? User wrote English with some tags. Need follow prompt engineering instructions. Need likely output

27B dense today beats opus from half a year ago.
27B dense will beat fable by the end of 2026.
Trust the plan.

>>109555428
Just wait some one will compress it down to 16gb

File: 1769667249490796.webm (2.99 MB, 1280x720)

2.99 MB WEBM

>something actually unironically happened that's on topic

>>109555448
caveman speak = distilled from gpt

Hmm, I need a heretic version of the new qwen, it doesn't even tell me how to grow psilocybe cubensis mushrooms

>>109554476
I'm begging you anon share your workflows. I wanna gen more gemmys.

>>109555448
>Need follow prompt engineering instructions. Need likely output
it's j-space? "we kindly need to do the needful, user must not know us ESL. output good? yes, need respond now."

>>109555448
They recommend only setting 262144 tokens for reasoning. It needs to save on tokens. It's so efficient.

File: Screen_20260814_093140_0001.jpg (658 KB, 1222x1559)

658 KB JPG

using gemma's system prompt on qwen 3.8 27b q8

>>109555448
>chinkified engrish?
I think that's a desperate attempt at compressing reasoning traces.

How censored is xhe

File: crown.jfif.jpg (28 KB, 554x554)

28 KB JPG

Holy shit this thing likes to think a lot.

>>109555402
are you on unslop? try the recommended preset

But can it mesugaki?

Been testing out muse glimmer for generic storywriting. One, it's devastatingly retarded. Two, it reasons way too much and compounds its retardation in said reasoning. It does grug caveman style reasoning and almost verbatim repeats the entire user message in it and makes blatant mistakes like "the instructions say continue in third person. But before it was first person?" when there was zero first person writing and at most, some system instructions could've been construed as second person. It also acknowledged a 0 depth instruction to use a tool for persistent memory but decided to not use said tool because it was "mid session" at two rounds of chat (it should've used the tool on the first message).
What kind of braindead shit is this: "Also need to register memories? The system says at startup activating OptMem mandatory run memo wake. We didn't. Might be okay? Might need to do it. The instruction says at startup activating OptMem mandatory run ~/tools/memo wake before any other tool call, at first message of a session. We are mid session? Might be okay.
Probably just continue story."

The only saving grace to it so far is 1) relatively cheap context 2) sort of creative, in this sci-fi story it actually came up with somewhat interesting creatures and names for native flora/the planet. It's ass in all other ways though.
3/10 model. Would prefer to use granite 30b over this

>>109555472
Also helps shitify writing as that's an anti use case that hurts inference providers due to lower mtp match rate.

So now that the dust has settled, is Qwen 3.8-27b a resounding success or big old flop?

>>109555490
Yes

>>109555471
What is a fake word? If it's a fake word how does a LLM know it in the first place? It's more like a zen riddle.

>>109555486
>Would prefer to use granite 30b over this
that's fucking grim

>>109555490
>settled
I aint even finish download yet..

>>109555486
Zucc bros...

>>109555486
>almost verbatim repeats the entire user message
Ya I noticed that too. Very odd

>>109555490
unironically give it 2 weeks for the jeets to calm down, then we'll get the real data. I have a feeling people will be going back to 3.6 for some things

>>109555459
It's simply the default reference-to-video workflow from ComfyUI.org with three reference images of the various Gemmas and several attempts before getting a good gen with a non-autistic prompt.

><Subject 1> is the character referenced in <Picture 1>.
><Subject 2> is the character referenced in <Picture 2>.
><Subject 3> is the character referenced in <Picture 3>.
>They are all are 11-year-old mesugaki girls.
>
>[Shot 1] The scene is an empty Japanese-style elementary school classroom. Closeup view on the girls, from the side (we can see their faces clearly).
><Subject 1>, <Subject 2> and <Subject 3> are all standing in front of the blackboard, scribbling obscene drawings on it, laughing and having fun.
>The three girls are not wearing a backpack.

>>109555481
what preset? i have it at
Thinking Mode: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0

>>109555223
???? he puts out gguf, you need gguf instead of

.doesn'tworkonanything

>>109555486
I tried using glimmer on claude
its as retarded as qwen 3.6 it needs decent amount of handholding
that said its internal reasoning style is a little different

>qwen3.8-27b has a deepswe1.1 score basically the same as glm5.2

there's no way this shit isn't ultrabenchmaxxed
has anyone used it for real coding yet? i refuse to believe that a 27b model can actually do well at that benchmark, let alone be able to even call tools reliably

>>109555519
I even managed to forget to delete a word in the prompt that I didn't notice when I generated the video.

>>109555532
>has anyone used it for real coding yet?
it's been minutes bro calm down

>>109555486
>can't load sm tensor on llama.cpp
shan't be using it

>>109555500
Granite is in that "he's a little confused, but he's got the spirit" tier of models, like jamba. They have no fucking clue what they're up to or doing, but they try and sometimes succeed at something

File: q.jpg (134 KB, 1115x628)

134 KB JPG

sqeeeeee!

>>109555532
It's still a dense model and maybe we've been led to believe tiny models can't be good for anything considering there's a multi-trillion financial incentive to say that

>>109555532
its literally out there now why not try it yourself

>>109555486
promptlet and harness issue
glimmer requires specific wording in system prompt
just like you have to use the word “policy” to uncensored and not any other words
system prompt should be model specific, generic harness is a mistake

>>109555532
Maybe they used that fable data anthropic cried about?

>>109555556
It works, but you have to edit a line and recompile.

>>109555532
You're tripping bro Alibaba clearly got four times better at AI.

There's no way they didn't train the MoE alongside this. Once this initial hype spike dies down, they'll tease something else that's coming to get the jeets shitting in the streets again

WHY IS THE DOWNLOAD SO SLOW

File: balls.png (174 KB, 1640x1058)

174 KB PNG

>>109555471
Cute

File: file.png (102 KB, 1235x931)

102 KB PNG

it thinks a lot

I will wait for AA and calculate non coding non agentic score from their run. 3.6 27B is lower in this score than 3.5.

>>109555583
Llama.cpp stops doing anything and I can't control-c out of it when I send an image. I'm also using a gemma 4 31b sm tensor patch, maybe that's affecting it?

File: glimma.png (7 KB, 815x78)

7 KB PNG

>meanwhile I'm still downloading glimmer because I've been too busy to test it

>>109555604
nice, prompt?

>>109555471
Reads like an ironic edgelord making fun of weebs while being a weeb himself.

>>109555604
hey, if qwen is that good at writing the- oh it's gemma.

>4 minutes for thinking
bruh wtf is this trash ass unc model JSID fr

Ask your Gemmers (or Qwen, if you're some sort of fruity shotacon), to complete this pattern - JD, JK, JC, JS, J?. It's fun to watch her flounder.

>>109555606
>thinks a lot
>unsloth
>IQ2_XXS
no anon that's just brain damage

you guys did read the model card right? you know it defaults to xhigh thinking and you can turn it down to medium/low right?
right?

>>109555566
brother if you have an immediate message saying "use the shell to run this python tool with these exact commands at the start of any chat" immediately before the user message and it simply decides not to, I don't know what sort of harness or prompting you think is needed
>just like you have to use the word “policy” to uncensored and not any other words
I mistakenly assumed you had some knowledge until I read this butchered line of poo text

uhh no i just asked chatgpt to give me a ollama cmdline to the best model

>>109555559
>friend
Already friendzoned, NGMI

Are you really gonna cheat on Gemma with someone called QWEN?

>>109555655
stuf

>>109555653
I sometimes feel like I'm the only person on /lmg/ who actually reads the entire model card. This place has gone to utter shit and is filled with so much tribal shitflinging retards.

>>109555604
Oh hey it's Princess Gemma

Is Qwen a male or female

Is Unsloth really that bad?

>>109555645
mine cheated cuz it looked online :(

>>109555668
Trans. Original name was Dan but now she calls herself Qwen (Gwen was for normies)

my balls are vibrating because i havent cummed in 3 days

>>109555666
If they didn't fill it with so much slop generated by the model literally praising how it's the best thing ever maybe more would read it

>>109555668
Asian ESL autistic tomboy
>>109555669
98% of the time yes, but occasionally has god-tier specific quants

>>109555666
do you expect a general that is dedicated to LLMs outputting mostly text would actually read text?

>>109555669
It's a dice roll. Sometimes their quants are okay, sometimes they're worse than they should be for some unfathomable reason. I used their GLM5.1 quant for months didn't see a difference to the API but their exact same quant for GLM5.2 was fucking shit.
You'll also have to be prepared to deal with chat template issues.

Okay so Gwen 3.8 reasoning effort is a big deal. Prompt: "Tell me a children's story featuring anthropomorphic animals, roughly 300 words long."

Default xhigh reasoning: summarizes requirements in grugspeak, drafts story in CoT, proceeds to restate the draft with numbers after each word to carefully count them, grows concerned that the total is only 250 words, carefully plans out a series of edits which will bring it up to exactly 298 words, dithers about whether the title is included in the word count for a bit, restates the entire final draft in CoT, then responds with the story.

Medium reasoning: "Tell me a children's story featuring anthropomorphic animals, roughly 300 words long.", then proceeds to write it without thinking.

>>109555686
now imagine if llama.cpp had an easy way to adjust this without doing kwargs and reloading the model

File: 1780865254133157.gif (1.35 MB, 342x316)

1.35 MB GIF

>>109555666

Sure we read it, but I'm pretty sure that everyone here has the basic expectation that the thinking would be at least functional out of the box.
Not so completely broken that the model thinks +5 minutes on every question that's more complex than "What's the tallest dog breed" and you need to tune the default down to a level where it actually works.
If their default thinking is borderline broken, they should point that out at the very beginning of the fucking model card.

>>109555686
And how many words was it on medium, how off was it?

>>109555664
double stuffed oreos?

You wouldn't fuck a capybara

>>109555655
because you’re unironically a promptlet
glimmer has never failed to call mandatory tools in my harness

qwen 3.8 too much of a prude for me, going back go gemma

>>109555686
Whoops copy-paste error, the medium reasoning text was supposed to read "The user just wants a children's story. Let's write a story of about 300 words."

>>109555696
260 words, which feels like it's an acceptable interpretation of "roughly 300" to me.

>>109555703
you seem upset
Is this really worth your hourly wage?
Does honest opinions really activate your almonds this much?

>>109555692
you shouldn't have to reload the model, you can attach the kwargs to the request (assuming your frontend allows you to at least)
but it would be nice if llama.cpp did this in a way that was compatible with the API of every other cloud provider instead of through chat template kwargs

File: kwargs.png (132 KB, 1049x444)

132 KB PNG

>>109555692
Something like picrel? It'd be crazy.

>>109555677
I meant I read all the technical relevant stuff and scroll past the slop obviously. Every model is different and it's baffling how many people shit on models that aren't configured correctly.
>>109555694
Or just read and set it properly first time with no issue? It's obvious they purposely set the default params to whatever will get the best twitter demo results because they know most indians wouldn't think to increase thinking to max when creating shiny demos. First impressions are everything and they hyped this up for so long.

>>109555669
You probably should wait as they will update their quants 10 times before having something good.

>>109555724
>260 words, which feels like it's an acceptable interpretation of "roughly 300" to me.
Agreed.

abliterated-liberated-uncencored-ccpremoved-bratty version when?

File: Screenshot 2026-08-14 at (...).png (218 KB, 1678x1042)

218 KB PNG

Looks like Dipsy Flash 0731 beats Qwen 3.8 in almost every benchmark, and it's going to be 3x slower on unified memory systems. And that#s beside not having a speculator model yet

I think I'll stick to 0731, but happy for you 32 GB VRAM havers.

Any J-Spaces experts here, is QWEN female coded or male coded