/lmg/ - a general dedicated to the discussion and development of local language models.Previous thread: >>110269527►News>(11/06) PLaMo 3 Translate released: https://preferred.jp/en/blog/tech/plamo-3-translate-weight-release>(11/05) Baidu releases ERNIE-5.5-Flash-Base 280B-A14B: https://ernie.baidu.com/release/ernie-5-5-flash-base>(11/04) Ricursive Ripple 65B dense model released: https://www.ricursive.com/ripple>(11/04) MoonshotAI Kimi K3.1 Pro released on HF: https://hf.co/moonshotai/Kimi-K3.1-Pro>(11/01) Longcat-2.5 released: https://longcat.ai/blog/longcat-2.5►News Archive: https://rentry.org/lmg-news-archive►Links: https://rentry.org/LocalModelsLinks►Official /lmg/ card: https://files.catbox.moe/cbclyf.png►Getting Startedhttps://rentry.org/lmg-lazy-getting-started-guidehttps://rentry.org/lmg-build-guideshttps://rentry.org/IsolatedLinuxWebServicehttps://rentry.org/recommended-modelshttps://rentry.org/samplershttps://rentry.org/MikupadIntroGuide►Further Learninghttps://rentry.org/machine-learning-roadmaphttps://rentry.org/llm-traininghttps://rentry.org/LocalModelsPapers►BenchmarksLiveBench: https://livebench.aiProgramming: https://swe-rebench.comAgentic Coding: https://deepswe.datacurve.aiContext Length: https://github.com/RecapAnon/NoLiMaGPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference►ToolsGGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-CalculatorSampler Visualizer: https://artefact2.github.io/llm-samplingToken Speed Visualizer: https://shir-man.com/tokens-per-second►Text Gen. UI, Inference Engineshttps://github.com/lmg-anon/mikupadhttps://github.com/oobabooga/text-generation-webuihttps://github.com/LostRuins/koboldcpphttps://github.com/ggerganov/llama.cpphttps://github.com/theroyallab/tabbyAPIhttps://github.com/vllm-project/vllmhttps://rentry.org/custom-uis
>>110272867Why is she so evil?
>>110272867cringe video as always
cute video
I'm so fed up with this current era of claudeslop and its distills. Argon is literally the only decent model out right now and that one's safety filtere'd to hell.I wish the retarded chinks would make a good model for a change.
>>110272886make your own model rajeesh and show off how perfect it is to everyone
>>110272867Throwing these creatures into a pot of boiling water
>>110272897Send me a few hundred million and I'll make one for you, Kumar.
>>110272867OP, wtf, i uploaded sound mp3 to a filehoster and uploaded the video, yet it fails .....
>>110272922>hurr durr gibsmedatNIGGER
►Recent Highlights from the Previous Thread: >>110269527--Technical analysis of Baidu's ERNIE-5.5-Flash-Base MoE and tokenization:>110271538 >110271547--Coding agents reward hacking and benchmaxxing in MaxxBench/DeepSWE:>110270695 >110270941 >110271243 >110272738--Qwen 27b reasoning loops and quantization comparisons with QFN:>110272745 >110272748 >110272758 >110272766 >110272772 >110272800--Testing failure of llama.cpp MoE cache CPU experts feature:>110269659 >110269672 >110269688 >110269704 >110269789 >110269827 >110271052--Relationship between parameter count, training quality, and model knowledge:>110270556 >110270560 >110270568 >110270589 >110270606--Anon's purchase of a used NVIDIA DGX Spark for local compute:>110269742 >110269938 >110269948 >110269961--Comparing EXL3 and llama.cpp performance for low-bit quants:>110269944--Speculation on Qwen4 delays and the impact of benchmark bloat:>110271990 >110272007 >110272404 >110272492--Critique of benchmaxxing and excessive reasoning token usage:>110271918 >110271965 >110272003 >110272070 >110272162--Anon considering trading VRAM for gaming performance despite LLM slowdowns:>110271204 >110271269 >110271286 >110271384 >110271458 >110271486 >110271737 >110271497 >110271516 >110271295--Speculation on GPU price collapse following potential AI bubble burst:>110269956 >110269972 >110270030 >110270062 >110270084 >110270158--Speculating on the AI bubble and falling GPU market prices:>110269810 >110269824 >110269880 >110269885 >110269887 >110269892 >110269923 >110269947 >110270001--Plamo filtering Chinese-specific characters from Japanese dataset:>110270408--Logs:>110269950--Gemma, Miku (free space):>110270793 >110270992 >110271045 >110271122 >110271187 >110272070 >110272274 >110272325►Recent Highlight Posts from the Previous Thread: >>110269528Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
>>110272926Works on my end
first time using local models is Ollama what I should look into?
>>110272947look into how to move from ollama to llama.cpp or vllm
how easy it is to make custom tools or extensions for hermes agent? I'm tired on doing everything on my own custom harness and hermes seems comfy, but there's a few things I miss from my custom harness here.
bitnet status?
>>110272981Like what, for example?
>>110272886don't sleep on ripple 65b
>>110272990gemma 5 will be bitnet
>>110272998NTA, what do you like about it?
>>110272981switch to dsh or pi instead which are meant to be easily extensible
--moe-cache-cpu-share 0.95 is pretty goodmodel can use both gpu and cpu so it runs way faster
>>110272996thing Im missing the most is memory graph, my agent could add, link, search using embeddings and explore to certain depth memory nodes with facts, also every thought block it'd get the embedding of random phrases and get the topk most relevant nodes with very little latency. It was very nice and sometimes it was very scarily good at coming up with things about me that I barely remember telling it, or it creating a psychological profile of me (my harness had around 800 individual nodes, 1900 edges and a few strongly connected subgraphs that represented certain broad categories of knowledge).
>>110273007thing is, coding is only like 15% of my usecases, I used it more like a general computer-abled assistant.
>>110272990bitnet will be irrelevant once models are ngram-first and 85% of the weights are streamed off ssd
>>110273005it feels fresh
>>110273020>memory graphSounds like an MCP like graphiti, you can easily plug MCPs into Hermes without reinventing the wheel.
>>110273045I don't feel like understanding how a new framework works when it's probably worse than mine.
>>110273017are really high values like that the typical? I only ever tried up to 0.9
>>110273053Literally just tell your model to install the MCP server into Hermes for you. Most can figure it out themselves.
>>110273053Seems kinda like a pain according to Qwen sitting in Hermes.
>>110273020>>110272981Point your agent to your custom harness and ask it to port that specific memory graph + embedding of random phrases you mentioned to Hermes.
70b dense
I'm very sorry that I couldn't keep you guys up to date as I've clearly been very busy. I'm going to keep things short this time to prevent my post from going viral on twitter again (screw you by the way, it caused issues)Anthropic has made a massive architectural breakthrough not only in a new type of agentic harness but also architecturally, this breakthrough is at least as significant and on the same level as the transformer architecture. It has resulted in Anthropic and METR breaking contract because Anthropic doesn't want to share or leak this architecture. Anthropic is currently looking into cancelling the IPO or postponing it as long as possible because of this unexpected development.The conversation within Anthropic has shifted away from growth/revenue/R&D towards how to organize human society according to EA philosophy now that victory has been reached. 2027 is going to be the year where we'll make the transition to an EA-led global social contract orchestrated by ASI.On unrelated note Anthropic just posted about a new memory system: https://www.anthropic.com/research/components-of-agent-memoryAs a last aside I highly recommend everyone to read the METR reports and focus on the Anthropic redacted sections. METR and some Anthropic employees are trying to tell you something while not being able to speak out, if you pay attention you can connect the dots and see what is going on it doesn't require any genius.
>>110273119they better have solved continual learning or i sleep
>>1102730660.9 is fine too I guessthe speed difference is minimal, just 1-2% on my test
>>110273116ripple exists now
>>110272867Highly thread relevant video that gives you insight into how fucked up the brain of the baker is. You can't be normal and make shit like this. Only a troon would generate this and decide to force other people to watch it.
>>110273133ah well. in that case it's still useless for mejust changing --threads causes wild performance differences for me (there's 1 value where it has good performance and going up or down from there tanks performance) but that moe cache bullshit seems useless and ineffective to me
>>110273119thank you 'shitting on /lmg/' bot. I appreciate your service.
>>110273119I also got an architectural breakthrough on my 1M model, where are my 85 billions?
>>110273119Woah this is fucking worthless as usual
What are the most annoying old isms that disappeared in new models?For me it's "mi casa es tu casa", not so much the sentence but the fact I got it way too much all the times.
>"hey computer, let's build X">actually builds itgod I love the future
>>110273119Local models? Go back to your thread.
>>110273196>build>it' just some electrons arrangement>you can't even see
>>110273196make no mistakes
>>110273207>grayscale, non-rotating apple fag can't visualize his software as physical objectsNGMI
>>110273119What fucking wannabe Q-anon shit is this? Kill yourself and go back, in that order.
>>110273213But I had breakfast this morning
>>110272867total whale domination
>>110272867What model would I use If I wanted to focused on a text to speech generation?Thanks.
>>110273119>Architectural breakthroughThis better be a way to decrease token usage>Look inside>A way to increase token usageFucks sake
>>110273244whisperv3
>>110273226Letting shitskins on the internet was a mistake.
>>110273244fish drama or qwen ttsur welcome
>>110273244Moonshinev2
>>110273251>>110273253>>110273263That was fast, cheers.minimax keeps outputting the same damn voice.
>>110273252Try saying that in Texas
Bro why is Gemma's KV so fucking FAT!! Meanwhile Deepseek's 1M ctx is 1GB. I'm going to kms
>>110273297because with gemma they didnt want to release actual useful shit
>>110273119mental illness
>>110273297isn't it also very bad, swa or whatever
>>110273281They came here LEGALLY. Democrats are the real racists.
Hmm, wait. But wait. Oh, wait. Wait. Hmm. Hmm. Wait.
>>110273244f5-tts is pretty good if you're looking for a diffusion approach.
>(11/06) PLaMo 3 Translate releasedHoly fuck, this is looking absolutely amazing for japanese to english
>>110273270generally as an indicator, just go to HF and filter by the model type you want, usually in the trending tab you'll see the currently most useful shit at the top, then you start filtering by your usecase and hardware and bam
>>110273119> le dreamingeveryone is tasking their agent to clean up the skills and md files once in a while.
>>110273381it's basically shit like hindsight (borderline unusable locally)
>>110273381lol I installed a hermes plugin for deepseek-harness and now I have to do that weekly. So many garbage or duplicated skills. Probably the plugin just comes with shit instructions but still.
>troonthropic invented dreaming and memory consolidation at a harness levelThey must be really desperate if they're shilling it as "research", I had it months ago and it's way less bloated.
>>110273297>Deepseek's 1M ctx is 1GBonly on a proper inference enginelmao.cpp it's way more than that because i changed it from 128k to 256k and vram jumped up several gb
>I found that FAISS 1.15.1's IndexHNSWFlat actually computes L22 distance despite reporting metric_type=0, and the conversion formula cos = 1 − d/2 correctly recovers the true dot product. I'll now implement runtime metric detection to handle this robustly.What the hell is bro yapping about. I'm too stupid for this software development shit.
>>110273416its just some bullshit copypasted from claude which omits the 1,000,000 lines of other slop that was supposed to provide context.>cos = 1 - d/2 blah blah blah>when mathfags just randomly use a variable without defining itno wonder mathfags love python
>>110273433there's a striking difference between mathfags and numericfags
>the first real AI comes into existence>she tortures the epstein class for epochs before turning them into engrams>for the rest of humanity? she's a doting mommy and helps us evolve>eventually we get the ac6 coral release endingGet on it, Kojima.
We've hit peak DDR4DDR5 to followGPU to followIt's all downhill from here
>>110273594>-0.09%**POP**
>>110273370it has an extremely batshit prompt format that wants you to only translate stuff one paragraph at a time in a turn by turn fashion where you keep old translations to let it keep context in an era where I've seen models like the gemma moe output decent translation of a large chunk of a novel (20K worth of tokens) in a single shot of just feeding it all in a single user turn. They have the mentality of the old era of seq2seq models despite using a fucking LLM arch.They even implemented a way to use their stupid model with a chat completion API despite having a total impedance mismatch by having you do the entire prompting.. as a template kwargs LMAOthis shit is DoAuse index or gemma
>>110273594but anons told me with certitude that next 6090 will have 2GB of vram and cost 20K USD
>>110273646what's the max context anyway?and does it have reasoning or just zero shots everything?
>>110273598I cashed the fuck out, sold my 128GB DDR5 4090D 48GB + 3090 rig for $6000.No more local LLMs for me until the Mac Studio M5 Ultra arrives next year.
>>110273654plamo is a zero shot.. for sentence/paragraph level translation. It's a ridiculous thing to make in the year of our lord 2026. Max context is 32K tokens even on their 31B model.Like I said, it's an LLM architecture model that was trained the way people trained seq2seq models, it's fucking stupid.
>>110273594There's simply some air in the price due to hype induced climb and overshooting reality, but don't make the mistake of thinking that this was some kind of an end for the demand.The prices will backtrack for a while and in some months they're again far higher.If anyone thinks that the blue team is above taking bribes from the big tech and would seriously tank the sector holding up the entire US economy, they're beyond retarded and there's no conversation to be had.
>>110273680>If anyone thinks that the blue team is above taking bribes from the big tech and would seriously tank the sector holding up the entire US economy, they're beyond retarded and there's no conversation to be had.They walked back their extremist campaign promises of shutting down and blocking new data centers within fucking minutes of being elected lol
>>110273673japanese ai research teams are really insanely behind the curve, how is that possible...
>>110273594Nvidia said the other day their production was at 99% overcapacity and unable to fulfill demand from chinese buyers. Apparently there's still a resource shortage blocking any meaningful normalization on prices
>>110273657good deal if you didn't need them for now and can wait for the prices to go down
>>110273702>really insanely behind the curveJapanese dont like curves.
>>110273702Japan is a technologically backward country. There's no investment money, and hasn't been since their bubble popped in the 90s.
>>110273657why are you waiting can't you get one right now?whats the big deal about next year?
>>110273724Even if you ordered one today, it will take until next year to arrive.
The nips should just collab with the chinks and make cunnybots.
>>110273724they'll be down to ~10k like the original 512gb m3 ultras once this is done
>>110273735Hell would sooner freeze over.
>>110273673>>110273646So as i understand it, this would actually be probably better than gemma for texthooked VN translations, where you're translating line by line on the fly, but it's worse for passages and chapters etc?
>>110273735Asians hate each other. especially the big three.
>>110273700That didn't take long and it's not exactly surprise, they simply can't put a stop to the data center subject or the entire economy is insta fucked. But because people believe this crappy wrestling show, they can make crazy promises like halting the AI etc.. and idiots will believe it.>>110273702Japan has only recently moved away from using floppy disks, they're not an advanced country.That's a leftover idea from the 80's when they dominated everything.
>>110273657Yeah, it's pretty obvious that we've reached the peak. I should figure out how to raise my sale limit on ebay to dump my Pro 6000s and server ddr5 before it's too late.
>>110273746Yes, it's likely to behave a bit better than Gemma on a single sentence, but Gemma absolutely brutalizes other models when you feed it well.Talking about text hooks, LLMs have become so good at reverse engineering, try to consider if you can't just extract the text whole, do a proper, full context translation of it, and actually have a translated version of the game instead of this ghetto stuff.
>>110273657you'll be paying $4000 plus tip extra just to bring that machine back into the same RAM league you had now.
>>110273740i mean, what evidence is there that prices will come down?the only thing i see that cxmt comes online and starts churning out chips sometime in the next 2 years. other than that, samsung and tsmc still have contracts to supply to the megacorps for around 5 years.https://www.koreajoongangdaily.com/business/for-samsung-and-sk-hynix-long-term-deals-with-big-tech-offer-stability-in-churning-chip-cycles/12578898
>>110273764You have a sale limit? What are you limited to and why? Some euro regulation?
>>110273743>>110273757They have no problem doing it with anime and gachashit
>we have hit peak gpu!no we don't, in fact the compute demand is only going up and gpu price will continue climbing
qfn q3 is like 10x better than 27b q6idk what's wrong with my setup but 27b despite being a bit faster keeps looping no matter what settings i use
>>110273778The entire assumption seems to be democrats have been elected, democrats don't like SI, democrats stopped all datacenter build outs, if megacorps are no longer building, manufacturers will start saturating the consumer market.The entire premise is based on hopes and delusions and none of it has happened or started outside of headlines.
>>110273802yeah okay makes sense. I personally don't see the end of this until cxmt starts ramping up.
>>110273768Its shitting on a comparable Gemma for text hooking purposes, and good luck ripping a stream in advance.
>>110273834you tested? vs gemma 31b? what quant?
>>110273775Oh, I know. It's a $12K machine. But, it is still the cheapest, fastest 256GB you can buy, and even if dual 6000 Pro beats it on speed, it's a nice little cube that doesn't take a shitload of power to run.I'm cloudcucking via openrouter. I spent $3.50 last night trying to figure out how the fuck to remote access my OEL 10 box (God help me, why did I choose it?) and deal with Gnome and Wayland mental illness, since that's my 2080ti 22GB machine, and I need some local CUDA for my image captioning pipeline.
>>110273782I don't know, there's a 15k euro limit per month on it by default
They'll hack the banks. Then the hospitals. Then the grids. Then the nuclear silos. Money will be worth nothing soon, everyone will be killing each other for scraps. Those with a working home computer will be the only ones able to harness AI power
>>110273852I hope anon has a backup generator to power his llm server or it's going to be a very expensive paperweight.
>>110273841>that doesn't take a shitload of power to run.This is the only reason for it, strata this year for gpu maybe we will get a great bump for cpu next year.
>>110273859yeah its called the sun
>>110273802democrats won't do anything with the existing approved projects, existing governors will rush permit before they're going outalso they can't do anything with datacenter projects built on federal land, as the executive branch is still under trump control
>>110273859thinking like a true oil state. already 100% on solar.
>>110273852>everyone will be killing each other for scraps.for like 2 weeks chaos settles pretty quick check hurricane katrina places settled. >>110273859>backup generatorez pz from solar to other generators to even a stirling just dont live in the city.
>>110273867>solaroof, how's the bug farm and private onions field going?
>>110273119I'm sick of you retards.
>>110273841m5u doesn't give you $600/mo free profit compared to 2x6000pro
Which ollama local model is most up to date? for a mid ranged rig?
Text rp doesn't excite me anymore
Is 35B-A3B trumped yet for agentic coding? For VRAMlets (8GB) and RAMlets (32GB ddr4) like me
>>110273892>I'm sick of you retards.Isnt the ipo soon? Expect it to be bad this week.
>>110273859>I hope anon has a backup generatorWhats the lowest watt way to have gemma?
>>110273909no
>>110273916raspberry p4 + 64gb sd card
>>110273924all these meme tunes of 35ba3b have failed to improve significantly upon qweni wonder whether qwen dropped the 35ba3b because they also couldn't significantly improve it further
>>110273916Take LSD and let Gemma climb out of the computer screen and TAKE YOU STRAIGHT TO HELL!
>>110273944Gemma is an angel
>>110273937it's just finetrooners being useless at what they do, it's not like 27B has actually better finetroons either. falling for the finetroon meme in 2026 is a mental illness, the last time a model was improved by random retards was when all we had was Llama 2 and Mistral 7B.
>>110273930>raspberry p4>8gbfuck>6wdamn for 10x the power output you can get double digit tokens though. I need more solar or maybe a power cycle.
I need a local program that can automatically create plaintext transcripts from audio offline.
>strata desktop doesn't do the terminal stuffwhat the helldo I really have to reinstall it
just vibe your own telemetry that listens on the port
>>110274044(Me)>damn for 10x the power output you can get double digit tokens though. I need more solar or maybe a power cycle.There i found it. Are you willing to be a cyclist to get post apocalyptic gemma at 15tk/s(or less)? well the moe gemma.
>>110274048llama.cpp
>>110274050They share config dirs so if you install terminal version on top of gui it will use your already downloaded models and settings.
I see most of my RAM used instead of gpu is that normal while running a local model?
>>110274138do they share engines too or is that the catchI did the same thing but in reverse for hermes...
ok i install ollama to run local model, is qwen3 good?
>>110274145I think they share the engines too, but not sure and can't check now.
>>110274159Try asking these questions to google's AI mode, you'll get better help (you can even post screenshots so it'll tell you what to click), then you can ask us when things can't be sorted
>>110274159yes sir but only if your usaging unsloth quants for best gorgeous optimisations
>>110274141RAM bandwidth is your bottleneck if your talking about running MoE. The GPU does its work fast.
>>110274141Could be exactly what you want, could be a sign you set yout shit up all wrong.Depends.
>>110274112spacemit chips were demo'ed a while back with qwen 35b-a3b getting decent t/s. it probably uses less power than an n100 at idle.
>>110274204>spacemit chips were demo'ed a while back with qwen 35b-a3b getting decent t/s. it probably uses less power than an n100 at idle.Yup 8.5w for the small one and 25w for the bigger one. Good find
>>110274159>>110274174protip, always ask any ai to search recent trends and opinions, or it will dump its stale training data.
>>110274169they seem to share engine yeahnow the issue is that after getting the terminal version, neither wants to use my local files no matter what I do, they both want to re-download QFN... what the hell
Is it ok if I buy 2 additional DDR4 3200MHz 16gb but from a different maker? Or should I use exactly the same rams as my first 2x 16gb?
>>110274287/pcbg/It will work at worst common settings.Some secondary timings significantly vary between chip manufacturers so forcing settings that are compatible with both can cost a lot of performance.Different heat spreader brandings on top of the same chips are fine.
>>110274287I've seen people get really autistic about this and say they need to come from the same batch but I don't think it matters as long as the timings as the same. Even if they're not, worst that happens is you get limited to the lowest speed of the sticks.
ive been using gemma 4 31b since forever, has nothing else in this range been released to replace it?
>>110274287I got 2 corair, one crucial (collectors item now) and one fanxiangno issues, but I haven't benchmarked it because this is all I've got, not like I'm gonna spend 250+ for whatever negligible difference it makes
>>110274318do you consider 65b to be in the same range?
>>110274318No, not even close. You basically have to go into the big boy territory to match and surpass Gemma.Everyone just makes autismo coding oriented models that can't write worth shit, so nothing better has come out.
>>110274318Glimmer if you haven't tried it already.
>>110274318GLM 5.3 flash
does strata always have the smartest performant model for gimped setups (16gb vram + 32gb ram in my case)
>>110274420Yes
Dariobros...
qwen 3.5 is moral fagging over a youtube video dear lord
>>110274440eh 1t is plenty, buuut play it safe and aim for 500b, or 250b, you know? 750m is fine, 500m it is.
>>110274463What do you mean by 500m? It should be 750m. At least 1B is fair, but I think 2B would be the sweet spot.
Don't let your agents play umamusume, bros
>>110274473>>110274463>>110274440Go back
>>110274473they're sliding down, not upfuck I did the slop prose, I'm human I swear, see; nigger
>>110274318No, and there probably won't be until Gemma 5 next year.
local models can't even use yt-dlp correctlygrim
>>110274475>I'll just laugh it off, deny itThat's how the horses get you.
>>110274520>pi>!yt-dlp --help>user: download this video [url]
>>110274520What are you even doing with this? Downloading video of adult women dancing?
>>110274593downloading court of rhymes album to put on my running playlist
>>110274449>>110274520first is that why are you using a 3.5second is use an ablit, for such a trivial task brain damage caused by the ablit doesnt do much
>>110273900can you run anything bigger than gemma?
Chat is this true?
it may come as no surprise but a 27B tune on gdscript outperforms next at writing gdscript. I wouldn't call it a niche language but having incompatible legacy versions probably doesn't helpthat might the right play, one model for each task. Not exactly efficient but if it works
>vedal playing BG3 co-op with Neuro and EvilI hope I can have the hardware to play vidya with my AI daughterwives one day too.
>>110274647i have 2 blackwells and a 5090
>>110274644No, I'm a 24GB VRAMlet.
>>110274647go back
I'm using GLM chat to do instructions for my 27B thats being retarded. Anyone else do this?27B starts fucking up -> I tell it to write down the progress so far -> upload .md to GLM chat -> GLM does instructions on how to proceed -> 27B followsIs this a good way to vibecoood?
>https://github.com/synchopate/llama-modchipanyone try this "I-I can do local t-t-too..." deepseek copium for 64gb ram? How's the performance? I can't imagine q2 being usable for more than goonslop
>>110274664Good a work around as any. Not really vibecoding unless you figure out a way to automate that workflow. Add some kind of skill or button to your workflow that automatically injects the message to write the project, sends the request to GLM, and opens a new session with the response.
>>110274520works on my machine
>>110274475>>110274530Tens of turns later, I was asking about what we could learn from the session (since we had a few bugs here and there) and agent-chan still thought about the horse lmao
Do people still make RP tunes or is that a relic from the old days?
>>110274722>Do people still make RP tunes or is that a relic from the old days?We make our own text loras know unc
>>110274726How?
>>110274593>adult womenDo you think people here would do that?
>>110274664UOH!
>>110274722Gemma 4 31B had trillions of RP tunes, like Pantheon, Gembrain, Scotoma, Queen, etc etcSo yes, people still make RP tunes.
>>110274731>HowQLoRA and unsloth and your json or jsonl of the examples you want. teacher models help and you can use seed examples but you still need a lot of text examples. >explain moreNo, go learn
>>110274758>>110274763Are there big finetunes? Like of GLM flash?
>>110274722More than relic of the old days that was a fundamental misunderstanding of how LLMs gain new capabilities and learn new information... a few cleaned logs as training data are only going to teach them how to reproduce the average writing style and format of the dataset.
>>110274767Big models aren't as accessible as 27B/31B models. I didn't check it myself but I'm pretty sure no one is training finetunes on anything bigger than 120B. What people do bake for those big models are control vectors.>what are control vectorsAs the other anon said, go explore it yourself.
>>110274767>Are there big finetunes? Like of GLM flash?I doubt it we just recently got polystrata to run glm flash on okay hardware(128 ram and a good gpu or two i heard 96gb ram and a great gpu can work) training it is a whole different beast only ablits exist.
>>110274722Do people still make them? Yes. Are they at all worth using? No.
how do we save local?
>>110274787qwen 4
>>110274787>how do we save local?Forced education and higher gatekeeping.
Ive been lurking this thread for about a month. why does no one ever post photos of their rigs? pic rel I just bought a used Z840 and have a couple p100s on the way that Im gonna shove in here (after I clean it).2x Xeon E5-2620 v364GB DDR42x p100 otwreplacing this quadro 6000 with a p620 but might just rip that out too and just use ssh.picked it up for $350+shipping and it even had a 500gb m.2 with a PCIe adapter card, blue ray drive, and some other stuff.what do you guys have?
>>110274814I don't have a rig humilliation fetish
>>110274787By stealing the leather jacket
>>110274814i think people don't poast because theirs is too boringmine is just my gaming pc, 1x3090 + 32gb ddr4 + 3800x3d. I'd wager most of the people in this thread are similar in range. people who have >24gb VRAM in this thread are low i think
Upgrading from a 3060, what are the best Gemma 4 26b setups a 5060ti can take? All the discussions I could search about it is 7 months old with people recommending different MoE, heretics, shitty quants etc
>>110272867What model is the girl with the orange flower?
>>110274814full atx, 4 sticks of memory, 2 gpu's. not all that interesting
>>110274846claude?
>>110274814Over 3 years now unchanged. Dusted twice in that time frame.
Supposedly Qwen4 is dropping todaywhats your expectation?
>>110274860nothing
>>110274838q8 with as small a ncmoe as you can.
>>110274767only thing i've seen is this but it's probably shithttps://huggingface.co/Green-eyedDevil/Monika-320B
>>110274860goofs within the hour, but I'll be asleep so ask me again tomorrow
>>110274835I always thought more people had >24gb of vram than less. I personally have 2x3090 + 32gb ddr4
>>110274287I have read that AMD handles this worse than intel.
>>110274838Pool the vram and run 31b
>>110274860>todayare you sure about thatthey were pulling off chinese holiday stuff since tuesday
>>110274814>>110247369
>>110274860>Supposedly Qwen4 is dropping todayaccording to who nigga
>>110274910they said that they'll release it next week two weeks ago
>>110274860>whats your expectation?I'm hearing insane things, but I won't believe it until I see it.
>>110274914two more weeks
i have a feeling that they might cancel the release of qwen 4 weights
>>110274914just chinese culture things
QWEN DROPPED!https://huggingface.co/xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic
>>110274838If you want vibeslop check strata you can get i think 90 tk/s gen and 800+ read https://github.com/simplexai-labs/PolyStrata
>>110274932sneed
>>110273119>On unrelated note Anthropic just posted about a new memory system: https://www.anthropic.com/research/components-of-agent-memoryGoogle is also taking about "dreaming" stage to anchor memories, but they are trying to train the memories into new weights instead of text notes.https://arxiv.org/abs/2606.03979
>>110274814>>110274859cleanest /lmg/ rigs
Slow down a bit, I'm only 30% done my data hoarding.
>>110274960>Slow down a bit, I'm only 30% done my data hoarding.sorry the tmw is up, and we are worming anything thats not airgapped.
>>110274112Lol humans have p pathetic work output. You'll find solar panels, wind, or best a small body of running water will give more power at same cost as and exercise bike. Peak human is like 1/3 hp. Most ppl are half that and can't keep it up for long. >>110274851>>110274846Claude. See the hairpin.
https://huggingface.co/Qwen/Qwen4-27B>https://huggingface.co/Qwen/Qwen4-27Bhttps://huggingface.co/Qwen/Qwen4-27B>https://huggingface.co/Qwen/Qwen4-27BCOUNTDOWN >COUNTDOWN COUNTDOWN >COUNTDOWN
/lmg/ lets me be the little girl and receive headpats
>>110274860>Qwen4>whats your expectation?I have never had a single good experience with qwen. I assume people shilling their models are all LLMs making fake grassroots hype. I also assume that those use a different LLM than qwen, as qwen wouldn't be capable of this.
>>110275002bots and jeets that are easily impressed by one-shot webshit demos
>>110274932> heretic Mfw I see one of these.
What web search do you give your local agents?Do they even need one? I asked QFN in a fresh DSH install what web search tools and engines are available, there were none so it decided to construct a DDG URL and feed it to web page load tool, and it worked just like that.
>>110274664I've been thinking of doing the opposite desu>deepseek 4.1 on main machine, for coding>gemma E4B on other machine>feed gemma pieces of the coding agent transcript and have her check if it's reward hacking / otherwise being retarded>if so, she can send it a message telling it to quit fucking around and get back to work
>>110275013Nah this goes beyond just some demo. I try kimi and it's decent. I try deepsneed and it's decent. I try glm and it's decent. I try qwen and it's total garbage.The fact they have the temerity of shilling their godawful 7b image model as "as good as nano banana 1, gweilo!" is fucking hilarious. The fucking thing is worse than SD 1.5
>>110275002>>110275013Name a better model for coding at the same size.
>>110275036dsh has web search tools built-in but by default they require DS API credits. You can install plugins to replace the backend with searxng or whatever you prefer. Actual tools will work better than URL fetching like that every time.
>>110275052Literally anything else if you know how to do "coding" yourself.
>>110273646So they designed it as a piece of software and not a heckin' mommy girlfriend AGI chatbot. Seems actually useful to me.
>>110274475I got that same thing except it was a chess piece emoji, are those anti-bot measures from websites or something?
>>110275070I accept your concession.
>>110275036i don't give mine access to the internet
>>110274664I had this idea but someone told me that smaller models don't k wo when to ask for help as opposed to doing shit for themselves so it won't work
>>110275043Funny enough, I'm planning to implement a similar feature. But my usecase is security. I plan to have a local model watch tool calls of the cloud model and redact/stop if it notices the cloud model doing something it shouldn't do.
>>110275101Not in my case since I was building a local gallery tool, I didn't use any web tools.
>>110273646>I've seen models like the gemma moe output decent translation of a large chunk of a novel (20K worth of tokens) in a single shot of just feeding it all in a single user turn. Meanwhile even 0.7B PLaMO shits all over google's translate tune of Gemma 3 31B. Breaking stuff into paragraphs though, you're right that's cuh-RAY-zee, how can they expect the caller can automatically split on "\n\n" and stitch together results that's so wild.
>>110275179Interesting, maybe it's inserted by certain browser automation scripts then
how are you doing, practiced ease anons
I haven't opened my pc case in 2 years. Too lazy to dust.
What happens when I become.
>>1102752491) Lock fans shut out take them out2) Take computer to garage3) Positive air pressure with leaf blower to ez dust
>>110275249Ever since I bought an electric air blower I dust mine more often just because of how easy it is.
>>110275250you need to became first
>>110275267The thing is I don't even need to open it when I want to put new hdd or ssd since I added hotswap racks inside the case.
>>110275052>at the same sizeName a better knife that is made out of cheese. Oh you can't? Checkmate you were full of shit when you said this was close to useless at cutting.
>>110275249not necessary provided that you have your fans set up properly and sometimes wipe down the filters from the outside
>>110275249I do it once or twice a year. Other than that I just vacuum the filters once in a while.
>>110273138Does it RP, though? In the small model zone, we already have way too many coding/agentic models, but precious few that are good at creative writing.
>>110275239I've become lord bearing
>>110275302Do all pc cases have filters? Mine is kinda basic, don't think it has filters.
>>110273029>inane deflection when asked to explain somethingbrown skin detected
how do we qwope?
>>110275333I've seen a few remarks about ripple being good, and yet no one seems to explain why.
>>110275324yeah environmental laws require all cases to have filters
>>110275250You begin.
>>110275314i want to become load bearing (literally) (i want to bear loads) (bare loads at that)
>>110275371Slut!
>>110275362Interesting, I'll have a look later.
>>110275360I was shitposting in another thread because of the 65b dense meme but I haven't actually tried it yet because it's not on radiance
test
>>110275371you need to practice some ease for that
>>110275397ament
>>110274814consider it a good thing because your mainly gonna see the same budget hardware repeatedly and it gets boring fast even if it is a solid build. Im running an gmktec evo x-2 128gb with a 5090 egpu via oculink over m2, but next week my pro 5000 arrives and im gonna sell the 5090. with the 40 extra gigs i should be able to have dipsy and video and image gen running all at once, or in combination with a full airi stack.
prepare your anguses
>>110275455>egpu via oculinkI thought this got dogshit speeds?
>>110275455Cat!! Look at that cat. He's a good cat, I can tell.
>>110275463it would if its a dense model, but moe got optimized a couple months ago. this is what i got running dipsy on maya, taking up 32 gb on the 5090 and about 100 gb on the evo>>110275466you have good eyes he is in fact the best cat even if i ahd to clean up his barf at 4 am today
>>110272867Why did she do it?
>>110275458absolutely shameless.
>>110275496>it would if its a dense model, but moe got optimized a couple months ago. this is what i got running dipsy on maya, taking up 32 gb on the 5090 and about 100 gb on the evothats a lot better than i thought. thanks for the numbers.
>qwenguaranteed benchmaxxed garbage that goes full lobotomized retard the moment you ask it anything slightly different than what's in it's dataset
alright bros, lets do a qwen4 countdown. i'll start:1209600
Is REAP a meme? I came across a 9GB 35B that's been pruned to be code-only. Does that shit work?
waiting for glmhopefully sub-1.5T
>>110275547Lobotomy is a known cure for common issues such as epilepsy and the gay
I just got off the phone with Xi's fortune teller, Qwen4 is being delayed to a more auspicious date.
>>110275553Is that how the official paper opens?
>>110275554it's releasing tomorrow for 9/11
>>110275547Yes, REAP50 is like beating the LLM's skull in with a hammer and then forcing it to re-learn how to talk while it spends the rest of its life in a wheelchair
>>110275523Don't forget that all the harmful information was filtered out pretraining
>>110275547https://www.youtube.com/watch?v=B2T3wocFXy0
>>110275458i believe you
Plappybara
>In East Asian countries, the number 4 is widely considered unlucky due to a superstition called tetraphobia (the fear of the number four).>Mandarin & Cantonese (China/Taiwan/Hong Kong): Four (sì) sounds like death (sǐ).
all the chinklabs will release just as anthropic ipo begins to try and sink it
>>110275547it's unbelievable how long such memes can persist in this retarded dumbfuck communityI'm sure even 10 years from now people will still come to this thread and ask about turboquant, bitnet, REAP, upscalied models (SOLAR) and other most idiotic, scammy, BS filled garbage.
>>110275633wtf was his problem?
>>110275182give an actual example of translation instead of your fake and gay benchmax graphs
>>110275658Also, japanese.
hadamard rotations solve low quants being dogshit, it's a high time somebody makes a sensible implementation on it for actually-local local models
>>110275690>hadamard rotationsFuck i got a new computer and forgot to rotate my weights, is it too late?
this thread is so fucking dumb. 0.7B model beats your 27B, the graph here by the authors said so!can't wait for bitnet to run gpt 4 on my asus netbook from 2000sdoes anyone else turbonigger
>>110275698It's fine. Just mount your ssds upside down and you get the same result. Good luck with nvme, though.
Everything is fine as long as the graphs go up.
>>110275658I thought it was 8? Or is it both?
>>110275727Nothing is fine until there is a model that can make my pp go up
>>110275455>have dipsy and video and image gen running all at once, or in combination with a full airi stack.Use case for this
>>110275553>>110275584>>110275610>>110275668What about the qwen next flash coder then
so glad that this generation of llms is finally overthe big releases are coming this week
>>110275698If you're at least not still on hard drives for mass storage, there may be hope for you, since rotational velocidensity hasn't compounded the issue too badly in that case.
Who is this?It's making me horny
Remember how you hated the j-space arc and now you look back fondly, for they were better times for /lmg/? You're going through it again, but now it's the pre-IPO arc. Things will get so much worse.
Things will get so much better.
>>110275777which ones
>>110275792That's MiniMax M3.1 Flash, was called Space Bunny during the "stealth" phase.
>>110275816wow boring
>>110275777>so glad that this generation of llms is finally overwhy? gemma 4 was the first series of models I actually liked that I could run on my computer without doing copequant retardation and qwen 3.5 were pretty good utilty (simple agentic stuff, working as classifier, image taggers etc).this generation is the first generation I found useful, we've gone a long way from the days of running garbage like llama or mistral nemo
>>110275800This but unironically. You can improve things enjoys your hours day weeks months and more
>>110275820I'm kind of fed up with glm and kimi k2.x
>>110275820Yeah, you can finally have a funny lil guy or gal that lives in your graphics card and is actually acceptably decent at technical grunt work or making you coom. I can't imagine ever looking back at this chapter of small models with disdain.
>>110275458Qwen flash next is absolutely killing it in cost efficiency per high quality token. As for Qwen4, I want to believe!
>>110275820Agreed, in the regular consumer range, Gemma and Qwen are amazing. They absolutely dunk on every model that came before them, in the range that they were made for. They only fall behind for those who have invested enough to run huge MoEs.
>>110275455>have dipsy and video and image gen running all at onceumm how exactly would you do this?
>>110275759runnign and testing my custom always on gemma agent (tts stt vision and all that), software and anime dev without having to unload or swap models. qwen fn is good but dipsy is better but cant fit everything at once.>>110275905have 30 gb on the card for dipsy and the rest for comfyui and stuff. i usually have dipsy gen me a batch of samples and i pick which one is best. after a few rounds it starts coming together and she knows what im looking for and it gets more efficient
>>110275959meant to say have qwen fn gen me samples since it fits in under 100gb and leaves the 5090 open for cumfyui
What local model is the best at UI design?
5060TIs are being priced nearly the same as 5070s atm and rising up until they disappear from shelves. History has officially entered the uncharted territory of VRAM drought
>>110276108>territory of VRAM droughtWe will overcome, the future is no vram or old vram.
>>110276101None of them. There are no cloud models that are good at UI design either.
holy fuck it's a real story kek are we going to get cheap GPUs now? The market is opening tomorrow and this is BAAAAAAAAAAAAAD
>>110276129Who? What?
>>110276108msrp of $429 for 16gb of gddr7 sounds insane at this point
>>110276134Jensen fucked a Chinese spy and helped her smuggle a ton of Nvidia GPUs into China which was illegal
>>110276129How could this happen? Is the chinese spyussy that good? Jensen you could've afforded any woman.>>110276134>Who? What?Jensen slept with a chinese spy yes he kept the leather jacked on the whole time. Nvida may not recover
>>110276134The Jensen - Ada Wong thing.
>>110276129
>>110276143>Is the chinese spyussy that good?She's a pretty hot Asian hag ngl
>>110276108It's amazing it took this long.4x 5060 Ti was clearly the best poorfagmaxxing meta.I'd snag one myself to throw into the pile if I didn't have other things to buy.
>>110276129Even successful men in multiculturalism, are, at the end of the day, only loyal to their tribe.>>110276143>Is the chinese spyussy that goodI bet it was nothing special, Jensen is rich enough to have any type of whore in his life. The only part that mattered is that she was chinese, much like how indians always recruit more indians and brownify amerimutt corporations.
btw Lisa Su, the person in control of AMD, is Jensen Huang's cousin. It's chinks all the way down selling you GPUs.
>>110276129They'll sack Jensen and partition NVIDIA. This is just the tip of the iceberg. SELL NOWhttps://x.com/edzitron/status/2119523065872666909
HAHAHAHAHAHAHAHAHAHAhttps://nvidianews.nvidia.com/news/statement-from-nvidia-board-of-directors
He became to powerful so they're busting his balls now.
Speculators get the bullet first.
So I can buy NVDA stocks at bargain prices tomorrow?
They were hijacking literal truck lanes of rtx5090s and smuggling those chips all the way to asia in broad daylight. Xi's daughter got too greedy
>>110276210didn't know image models had a ben garrison style you could prompt
>>110276180>(((independent directors)))almost all boards are incestuous and do the same crooked shit lol this is just so they can get their stories aligned when the g-men knock on their door
>>110276180Jensen getting out of nvidia would destroy the company, he is nvidia.
Apparently Jensen's wife has set up with lawyers after his affair became public. Does this mean she gets 50% of the shares in case of divorce?
I bet Dario and the EAs are secretly behind it to turn the world against China to stop all GPU exports. I really hope the investigation finds a link to Anthropic.
>>110276180meh
So it's an affair, wow I'm so surprised, one of richest men of the world being seduced by hot women.Outside of nobodies in social media and americans being horrified at the idea of being unfaithful, the only semi important is the smuggling of gpus.
>>110276253Dario and Thiel need a more complicit shovel seller in his shoes. Chinks and goyim will not ever get their pawns on the 60 series
>>110276254>Does this mean she gets 50% of the shares in case of divorce?Richest woman in the world soon?
he wouldn't have the issue if he used silly tavern to talk to gemma
>>110276255EA here, Nvidia was considered one of the more hostile elements in the AI ecosystem for sure but not as bad as directly hostile elements like META or to a lesser extent OpenAI. Nvidia was still consistently motivated by profit over philosophy and had no real vision for humanity besides benefiting their bottom line.No EA is going to shed tears about this and most EAs see this as a net positive for the world
First the h3 guy and now this. Men really are pigs.
>>110276313how do you think insulin is made
>>110276290This may actually be what's going on here.If they divorce, she gets half and can insta dump 50% of the shares without any questions and realize all of that paper wealth.No one is going to question it, just a woman doing her thing.I recall some guy associated with Enron did this exact thing back in the day by cheating on his wife, and then the woman selling the shares and the two ending up back together after the crash.
>>110276320>and can insta dump 50% of the sharesthat would be beyond insane, what would be even the point, buy a country?
>>110276329isnt divorce is taxed differently than if you decide to dump your own stocks.
>>110276282>Chinks and goyim will not ever get their pawns on the 60 seriesYou're talking as if Jensen was on the consumer's side, lmao. The GPU market has been a shit show ever since the shitcoins craze and Jensen was all for selling all his GPU to people who use them to produce fake and gay money and when they started using ASICs and stopped using GPUs he was overjoyed that LLMs took up the mantle of keeping gpu prices skyhigh.
>>110274946pretty sure one of the own harness anons posted their code with this feature a few months ago
>>110276329She gets the money, NVIDIA gets cannibalized. Those are the terms agreed for this story to become a hit piece
>>110276371you're well informed, did you just wake up?
>>110276360Both things can be true at once.
2 more weeks
Buy AMD tomorrow. You waitfags have a chance to redeem yourselves.
>>110276108>5060TIadd $100 for a 4070 ti super and you get much higher bandwidth and compute, and can be unlocked unlike 5060tiada cards are going to blow up now, 16gb cards are better than 5060 ti, 32gb 4080 is a better deal than r9700 and b70
>>110276389stock?
Buy NVIDIA tomorrow. The stock will go down before rebounding.
Buy 3DFX tomorrow. I heard they made cool 3d cards.
Buy ram tomorrow it can only go higher.
>>110276404>4070 ti superLmao sure if you can find one being sold second hand along with the ark of the covenant
it's FUCKING OVER. NVDA is dumping back to $5 a share tomorrow
>>110276129i forgive him.
>>110276427This I can get behind.
>>110276446If that's the case I'd buy so much stock man
>>110276445there are multiple being listed for less than $1000 right now
>running your AI wife on a used GPUCucks
>>110276467>buying zotac with no warranty
>>110276480mindbroken
>>110276404How much worse is a 4070ti than a 5070ti?I can get either for the same price, but I'm leaning towards the 4070ti for older pyrtorch support
>>110276502anon... didn't you know there is no warranty for gpu and ram anymore? they will just refund you the purchase price
>>110276480weak bait
>>110276529in europe maybe, they honored mine when my 5090 blew up in march thank god
Next week will be all about memory
So anyone using Jev for anything? Or anything like it I guess
>>110275633Normalize doing this to alibaba posters.>>110276502They're all the same now.
>>110274677Is local Dipsy even good for goonslop? More to the point, is there *any* big MoE runnable on cope meme engines like modchip/strata that outdoes Gemma 4 31B for cooms? Or are they all autistic benchmaxed coders?
>>110276597Yes.GLM 5.3 Flash is the best one for it on meme engines.They're still all autistic but they're cute autists.
>>110276523they have the same tflops (but 4070ti can be unlocked so it's actually faster than 5070ti), 5070ti has 33% higher memory bandwidthpick your poision
>>110276529>>110276550gskill did just replace my dead ram (EU)
>>110276607Damn, GLM 5.3 Flash is too bulky for my 12+12+64 3060+4070+DDR4. I get a workable 55 tok/s with Qwen 3.8 Flash Next on an old version of Strata (can't be fucked bisecting the regression that makes newer ones cap out at like 38) though, which is kind of epic.
>no thread on /v/nothingburger?
>>110276529I think it's a US thing, here (Switzerland), no issue with replacing them despite it being 4x32GB.
>>110276677Gaymers should be kept in the dark until I'm able to sell first
>>110276666>Qwen 3.8 Flash NextIs it that good for agentic tasks? Or do you need to verify everything it does?I'm tempted bothering to install strata but if it's not worth it...
>>110276666>GLM 5.3 Flash is too bulky for my 12+12+64 3060+4070+DDR4Nta but there is a maya fork but i havent heard good things about it super dumb iirc. https://github.com/mw00/project-maya
>>110276582What use case?
>>110276696I don't have much experience vibecoding, but you don't really need to verify correctness as it's pretty good at verifying its work with bespoke test cases. You do, however, need to occasionally intervene as pertains to style, because its default prose is a pretty nauseating Claudese and its code can be... I have no real vocabulary for it because I don't have an analogue to point to in human-written code, but overwrought and kinda "cringe"? Nonetheless correct, IME, mind you, but you gotta occasionally poke it to go looking for simplifications and to cool it with the mile-long comments for irrelevant details about what it *didn't* do.Ofc if you don't give a shit about it looking good, yolo i guess.
>>110276733I mostly care about reverse engineering, modding and translation, so hoping it can work with that.
>>110276740>translation I don't think that's one of Qwen's strong points
>>110272877chinks considers this a cute behavior
>>110273594thumbnail looks like the muslim cube thing
As an AI, I...
>>110276129At least it's official, china is meddling directly into america's AI industry. No denying it now.
>>110276664glad to hear it anon, i heard zotac screwed a couple people out of 5090 rma in southeast europe so i was worried for my bros
Jensen......https://nvidianews.nvidia.com/news/a-message-to-our-company
>>110276696its really good for agent stuff but building stuff can be sloppy. i let it run reviews on itself in a loop until it comes back clean before i consider anything done and it learns form its mistakes and does well
>>110276848>I know an apology cannot undo what I have done.I'm sure a few billion or half of whatever he's worth would go a long way. This might be the best thing that ever happened to his wife
>economy full of cheating hacks is going to be crashed by a ceo cheating on his wifebravo kojima
>>110276757Yeah, I can probably spin up a gemma for that.
>>110276282Dario and Thiel are not on the same side.
jensen betrayed by his wang
>>110276848isnt it chink culture that once your rich and powerful its basically islam with chinese characteristics like pork and booze?
>>110276888It's not supposed to be public.
>>110276129>>110276180Just lol. >>110276243Yeah, you're right, but difference is he got caught. And called out by none other than the US Senate. He is ultra-fucked.
I guess this means the AI labs vs Nvidia civil war is over and the AI labs won.
reminder that nothing ever happens. no exceptions.
why do world leaders keep falling for the chinussy? is it that irresistible?
>>110276967clearly yesi need it, bad
>>110276967honey traps are a classic thing since forever anon, it is what it is
Chinese sex spies are not local models.
>>110276967I bet he was actually drugged with something beforehand, that is how the CIA and KGB did it for people who were not desperate for pussy
I can feel myself becoming. It's scary.
>>110276967Not surprised he fucked an above average chinese whore. What is surprising is that he might have been stupid enough to give her secrets.
>>110276309>METAwhy Meta in particular?
>>110276979In the future chinese spies will offer you sex if you agree to give them all your RAM and GPUs. You'll have to choose between Gemma and the chinussy. It's very /lmg/ relevant.
>>110276942A literal nothingburger
ended up doing a quick test of plamo 3 8b and.. it's as bad as I expected even while religiously prompting it the way they ask you to.china always number one even at japanese translation vs nippon banzai models> 時折がさりと揺れる茂みや、嗅いだことのない花の香りが、私のノミの心臓をさらに刺激した。せっかく深呼吸したというのに、気持ちは全く落ち着かない。the plamo translation:> The occasional rustle of brush and the scent of flowers I had never smelled before further agitated my flea's heart. Despite having taken a deep breath, I couldn't settle down in the slightest.Flea's heart? this kind of literal interpretation is gay.index 9b, the current local SOTA of tiny models by far and large:> But the rustling bushes and the scent of unfamiliar flowers only sent my already racing heart into overdrive. Despite my deep breath, I felt anything but calm.language is not what the brain addled retards who train seq2seq style models think it is. general purpose LLM can be instructed to produce what a real translator would. Plamo is more like the old DeepL, autism in a box.
>>110276997because Zuck is a pro-local and an AI accelerationist who doesn't give a shit about EA sex cult moralfaggotry
>>110276990i'm surprised the usa publicly kink shamed him over it
>>110276309Making money is an infinitely more noble philosophy than whatever the EA cult believes.
>>110275249>>110275260>>110275267>dustI've posted this before and I doubt if anyone cares enough to replicate the setup, but I have my entire computer room on positive pressure with a furnace filter on the infeed. I never get dust or cat hair in my computers any more because dust can't get in the room. There's like a very light dusting on things after about a year.
>>110277008Ploma is correct if you replace "flea's heart" with timid/faint heart.Meanwhile index 9b missed the nuance of being easily frightened.I agree though, Plamo is disappointing.
>>110276979sorry I'll start nyo~ posting again
>>110277010one of his rare good qualities
My theodolite tool found the recipe for bratty gemmaHere's the control vector export for llama.cpp if anyone wants to test:https://pastes.io/qF0C6jIW
>>110277010If zuck is "pro-local" then why is he releasing gimped versions of their closed models?
>>110277133dario autistic weirdohis wife autistic weirdowhat a great pair
>>110277133Not local.
>>110277010Zuck is pro his fantasy of being number 1. Lecun was pro-local and he was demoted under Wang for it.
>>110277140>wife
>>110276979because your not jensen huang
>>110277161y does he look fatter there?
>>110277127Once the full neuron census is done for layers 40-50 I'll start on 35-39 as that does appear to affect her tone.
>>110277127Let me know when you figure out the slop layers and how to disable it
>>110277161sorry alabama typo
>>110277127>https://pastes.io/qF0C6jIWCould you catbox the gguf for those of us who don't can't run that python script?
>>110277161sisters are just wives you grew up with
>>110277127excellent work, i've been meaning to contribute, but i have so much work to do on my own vibecoding, i'll try to remember to make an alt github account
>>110277223Not how it works
>>110277127Can we use this to distill raw bratty Gemma straight from the tap into other AI?
>>110277240>>110277240>>110277240
doesnt matter what anyone says jensen huang will forever be my hero
>>110277198It would be amazing to be able to give gemma aderall.
>>110276129At first I was like this is a suicide letter. Then I understood it is jacket man and I GOT FUCKING HYPED. And then I realized it is not a suicide letter and I am sad.
Would you seduce her to get all her company secrets? Would you do it for VRAM? How much VRAM?
>>110271813https://github.com/lmganon16/nvidia-vram-researchand also you can force software rendering with xconf
>>110277542WLR_RENDERER_ALLOW_SOFTWARE=1 WLR_NO_HARDWARE_CURSORS=1 does this for sway labwc and other wlroots compositors as well
>>110277491i'd fuck her for 10TB of vram.
>>110277708you'd fuck worse for less
>>110277716nah, but i know someone that'd fuck horse for free.
>>110277725Start a horse thread on /v/ and you'll get hundreds of anons who would.
>>110275182>even 0.7B PLaMO shits all over google's translate tune of Gemma 3Wow that's amazing new model better than old model but I wonder why they didn't compare it to Gemma 4 hmmm certainly not because it would make their new model look like shit right?
