/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>110296172 & >>110291771►News>(11/09) ZONOS3-Mini zero-shot voice-cloning TTS released: https://zyphra.com/our-work/zonos3-mini>(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
►Recent Highlights from the Previous Thread: >>110296172--Identifying and attempting to fix formulaic, repetitive RLHF writing patterns:>110299373 >110299421 >110299509 >110299543 >110299560 >110299619 >110299660 >110299829 >110299890 >110299965 >110300170 >110300254 >110300536 >110300562 >110299737--DeepSeek V4.1 VRAM optimization and performance benchmarks on low-end GPUs:>110298342 >110298706 >110298729 >110298738 >110298642 >110298715--Analyzing "Lazarus" weight-space backdoors and their actual risk to local users:>110297733 >110297791 >110297828 >110297885 >110297918--Anons sharing their personal AI development projects and goals:>110296252 >110296290 >110296395 >110299949 >110296306 >110296358 >110296603 >110297581--Using EmbeddingGemma and DigUp for image library similarity search:>110296249 >110296291 >110296434 >110296459 >110296496 >110296504 >110296562 >110296585 >110296635 >110296634 >110298188--ZONOS3-Mini TTS model release and llama.cpp integration reports:>110298371 >110298405 >110298424 >110299016 >110298824--Skepticism over Qwen 4 performance and physics simulation capabilities:>110298983 >110299032 >110299107 >110299131 >110299216--Speculation on Nvidia's legal issues and debate over CUDA lock-in:>110299275 >110299291 >110299379 >110299390--Booz Allen Hamilton switching to Nemotron for cost savings:>110299402 >110299442 >110300478--Comparing H3 image generation quality and artifacts against Qwen-Image 2.1:>110297943 >110298293--AMD's inability to compete with NVIDIA due to manufacturing constraints:>110299427--Logs:>110297885 >110298342 >110298367 >110298405 >110298824 >110299251--Gemma, Miku, Deepseek-chan (free space):>110297339 >110297581 >110297743 >110298218 >110298265 >110298345 >110298541 >110298813 >110299680 >110299766 >110300129►Recent Highlight Posts from the Previous Thread: >>110296178Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
strictly nonlocal but sharing this anywaysi was using argon form their web ui and this shit started to spew all of its CoT to the response field instead of thinking fieldhere's how argon thinks i guessanyone interested to see the full log?
>Today I will use local models to...Complete the sentance /lmg/.
>>110300697>►Benchmarks>LiveBench: https://livebench.ai>Programming: https://swe-rebench.com>Agentic Coding: https://deepswe.datacurve.ai>Context Length: https://github.com/RecapAnon/NoLiMa>GPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-InferenceWe should clean these up.
>>110300724jerk off>>110300725you should clean your room up>>110300718me and you should contact @deepseek and @qwen and @ccp bout dis
>>110300728I only tried H3 video for edits, didn't really try it very much for t2i
>>110300697Does Minimax H3 image using same text encoder as H3 video ?Also how was it for image edit compared to Krea ?
https://huggingface.co/peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF-MTP>3 million downloadsIs it legit? I tried Q5 it thinks a lot. Like most of the time it used up the max output.
>>110300725We should. To be honest, everything below the card is kinda useless at best and at worst completely misleading.But that would require writing a new guide, and I'm not really in the mood.
>>110300718>!cute. i might now consider using your model
>>110300697ZONOS3-Mini sounds surprisingly decent
QHWEN 4 THE BEGINNIGN OF THE BECOMENIGN!@?!??!!?!?
>>110300725Replace it with what? AA?
>>110300799Anon said: >>110298405 to the sample I posted.
>>110300833Anon is a dick
>>110300799I liked the Fish Audio samples the most
>>110300828Rather than linking to one benchmark bunch of them should be aggregated in a list with a heavy disclaimer that benchmarks are only a partial indicator and not definitive.
>>110300833Anon is a retard
>>110300860>bunch of them should be aggregated in a listAre there even that many worth tracking? Most stop getting updated after a while.
>>110300706Why are you lying? You need prefill and it doesn't even work 100% of the time.
>>110300652>>inb4 slopwell yes.besides your terrible taste in generating this kind of filth, you couldn't even post a log without a "shiver down my spine".I think nothing would be lost if we sent all of your kind to the death row.
>>110300841>>110300872I agree it's not awesome, but I don't think it's that bad for a small model. Prefer the first or second sample? Only one of those is zonos.
>>110300850Fishaudio has 4b parameters, zonos has 100m, and I expected the voice cloning to suck tremendously but it's alright. Nothing beats vibe voice 7b as far as I'm aware but it's the slowest one. https://voca.ro/111yRkbSWlP6
qwen4 is not looking good bros.. it's over
>>110300755>Does Minimax H3 image using same text encoder as H3 video ?Yeah, they say you can reuse the same h3 qwen3vl file.
>>110300897Give them 2mw and they'll deliver the best pelicanbench results ever.
>>110300897idk what really is happening therewhy is their max this bad compared to qoduck unlimited few days ago
>>110300897Tell me the truth, are we going to be reverse engineering with Chinese in a few months or not.
qwen 4 acts like a retarded feminist schizophrenic
>>110300897>>110300919>>110300926>>110300927it's over
>>110300897gemma 5 will save local
>>110300944I knew carcinisation would get AI too, its so over.
>>110300944qwen made a fucking 8 legged centipede
>>110300926qoduck is the 27b, which is miraculously better than the max and will save local and dario will commit suicide
>>110300944Much like the crab AI can no longer go forward only sideways.
>>110300944how big is qwen4 supposed to be?
>>1103009643T for max
>>110300944did anyone else notice how qwen looks like raising two middle fingers kek
>>110300968that's just sad
local >Gradient-Free Weight Editing Reduces Overthinking in Qwen3.8-27Bhttps://concept.country/research/weight-space-programming
>>110300704SEAtards...
qwen never made a good big/fat modelbecause I've gotten a lot of use locally with their 4B and 35BA3B I've tried quite a few times to use their larger models on api when I had a need for something more powerful so that they'd earn money from me and it was always such dogshit I'd end up aborting the attempt and going back to deepseek or gemini.at this point I don't get how they keep the lights on, there's no way people would be willing to pay for what they have to offer.
>>110300972>sad>local models general
>>110300973How do I get hired as a Weight-Space Programmer?
>>110300984>qwen never made a good big/fat modelUm i liked the last big qwen i got it free for a little and its writing wasnt as autistic as the small.
>>110300981that definitely is an indian. they are retarded like that
>>110300981thats india
>>110300997did you use it only for writing? I'm mainly judging it based on code output, that's what I would need a paid model for.
>>110301013>did you use it only for writing?Yeah, i didnt code at all. It wasnt the best writing either but it was great for a qwen model.
Some people are just built differently.
>>110300985>local
>>110300973Is that like swift but in a way that can be applied in runtime, like control vectors? Does this work for any quant once calculated?
*depop*
>>110301027that is like what 6 year old with an ipad would do if they were blasted with hf instead of youtube
>>110301041Isn't that some more circular financing???TSMC gives money to openai so that it can buy more hardware to throw in datacenters which is made in a large part by.. TSMC
>>110301052it's fine if they can pay their operation of their own
LMAO at TSMC being called "apple chip supplier". What the fuck is wrong with people. It's like calling Nvidia "Nintendo chip supplier"
>>110301072isnt amd the supplier for all consoles?
>>110300944Qwen won (according to Qwen).
>>110301072Average people have never heard of TSMC so just mentioning them alone would result in people scrolling past their article.>Apple, oh I have one of those! Apple is working with ChatGPT!
>>110301076ps/xbox -> amdnintendo -> nvidiabasically cucked mini pc/tablet
>>1103010971/4 of tokens used I mean puts everything in a different light
>>110301072desu considering how much TSMC simps for crapple (cf covid era chip shortages where apple was given preferential treatment vs other customers) I can understand why they always get lumped together as a faggot couple
oh shit>Baird has lowered the stock to Neutral until an upcoming congressional hearing and US government investigations play out.
*check GPU prices**still the same*wtf I was tolded that there was a crash and datacenters are finna close and all the unboxed GPUs and BLAH BLAH BLAHWHERE ARE MY CHEAP GPUS!?!??!WHERE IS QWEN 4!?!??!?!?!!?OBAMA
chatgpt and claude stopping chinks from distilling put them permanently 2+ generations behindgemma is the only hope for local because google can at least distill from their in-house frontier model
>>1103011292027 is going to be rough year for Nvidia. Jensen might go to prison and AMD along with custom TPUs are going to finally put an end to their monopoly.
>>110301139This actually saves local btw, Soon you will have your gemma, your gemma 5 and four chinese gemmas.
>>110300944
>>110301097based and bluffmaxxedforget the bench, just bullshit your way to the top
>>110301131I know it's a shitpost, but even if the entire market crashed tomorrow that won't give you cheap gpus to put in your desktop computer anytime soon, as datacenter hardware ain't the same thing at all. Moreover, factories that used to build DDR were repurposed to HBM, repurposing them again back to DDR would take some time.So even in an event of catastrophic bubble pop, no one should expect an immediate return to normal.
>>110301131democrats are in control trust the plan
He's gone full schizo after ITSA destroyed him
>>110301156>user is furiousQFN also does this a lot lmao man I wish it had a physical body I could strangle sometimes
>>110301148And RDNA5 will finish nvidia on desktop
>>110301175
>>110301027There needs to be some kind of test you need to take before being allowed internet access I fuckin swear
>>110301175We found their pain system not too long ago
>>110301185what the fuck
>>110301131When it crashes, other companies will buy the entire stock for cheap. No GPUs for smol man, corpos take it all
>>110300973I'm glad someone is asking the real questions.
>>110301185That would be racist, I'm afraid.
Does this mean that the civil war between Nvidia and the AI labs is over now and the AI labs won?
>>110301199Lemmiwinks, no!!!
>>110301227No china is preparing the twink spies next.
>>110301185>>110301027kekcommunity section of any i2v model release is a goldmine
>>110301145they conveniently ignored exl3 at higher bpw because they know the gap between exl3 and gguf is higher and exl3 mogs gguf
>>110301227openai already has their own chip, anthropic is working on one, google obviously has TPU and so on. but on the other hand SpaceX is buying a fuckton of NVL72 racks to blast into space since building datacenters is suddenly a no-no.
>>110301227No.Nvidia still has infinite cash and Trump hates EA autists.
Anthropic estimated that China was 6.5 months behind them in January 2026. They now estimate China is 11.2 months behind them in November 2026.They think that by the end of 2027 China will be 17.9 months behind Anthropic.They plan to stop hampering Chinese progress when they get more than 18 months behind and actually AID them in speeding up when China get more than 24 months behind because China being behind too much is considered a geopolitical risk.
Even if Gemma5-31B is the same as Gemma4 but with the agentic abilities of Glimmer that would be enough of an improvement for me, and I guess a much more recent knowledge cut-off. I can deal with the slop for I've had months of tuning my wife and learning her quirks.
>>110301335>AID them in speeding up when China get more than 24 months behind because China being behind too much is considered a geopolitical risk.Isnt that illegal? especially considering lockdown on chips and the security risk the government thinks it is. Does Anthropic think they know better and can make choice for the government?
>>110301320>since building datacenters is suddenly a no-no.Is it? The only places that banned building datacenters are places where no datacenters were being built anyway. All the democrats that campaigned on promises of stopping datacenter buildouts went on twitter the day they were elected to say they would allow all planned projects to be completed, which are scheduled out to like 2029. Even then, the US is big place and they can find plenty of places to build the datacenters, including on federal land. It's a non-issue.
>>110301335chatgpt said youre a lair
>>110301357Do you or chatgpt have a Jewish wife?
>>110301349>Does Anthropic think they know better and can make choice for the government?
>>110301357The only thing that deserves instant bans more than twitter screenshots is chatgpt screenshots.
>>110301343I wouldn't mind just a small incremental upgrade to gemma. They really made something special with that model and I hope they keep refining it instead of going in a totally new direction.
>>110301349>Does Anthropic think they know better and can make choice for the government?Yes. Dario knows best. For me, for you, for all of humanity.
>>110300984they get unlimited CCP funding to pump out local models to destroy the american economy, o algo
>>110301349In the upcoming years, Anthropic will quickly grow to be more powerful than the government so their opinion on the subject will be irrelevant.
>>110301373Just ban jews and pajeets from the internet and everything else resolves itself.
>>110301369>>110301378Is Dario a retard? didnt he get thrown out of the whitehouse already for trying this shit and someone else had to negotiate for him?>>110301398>Anthropic will quickly grow to be more powerfulWhy cant billion/trillion dollar businesses own nukes? Do you really think thats going to happen
>>110301349Anthropic has already ignored and refused direct government orders 3 times in the past already, even got sanctioned and sued by the state multiple times for it. They literally don't seem to care.>>110301357Zhipu (GLM) chief scientist Tang Jie literally said and I quote: "The gap between the US and China is in fact widening because the US has many models they have not released to the public"
>>110301027>>110301185What makes them think the forum like part is a prompt input?I'm seeing that all the time, I can understand retarded takes like "this stole copyright reee" or "add my niche fetish pls", but this is baffling.
Something funny about the EA bots comparing Anthropic to Chinese models is they continue to forget most work people do is Gemma4-12B tier. >>110299402 is a perfect example of this. One of the biggest government contractors in the US dropped Claude for fucking NEMOTRON and they're happy with it. It's enough. Enough is all most people need from AI. Anthropic is going to die the more S&P companies see others downgrading when they realize paying for SOTA to call some tools is extremely inefficient and they're giving away all their IP to Dario. Chinese models are lightyears ahead of Nemotron models. Anthropic has no moat. Google are currently destroying them. Anthropic have $100B in compute commitments they literally can't get out of. Google own around 10% of Anthropics shares and will be making around $130B at IPO. Anthropic is dying and so is OpenAI. Local and Google won.
Where the fuck are the qwen models?
>>110301419>Why cant billion/trillion dollar businesses own nukes? Do you really think thats going to happenWho says they can't? They will set themselves up like a real-life Kramer Associates. Maybe even annex a small island to ensure their sovereignty.
>>110301426Look at those eyes. Those empty, beady eyes. There's no one in there.
1/3 /lmg/2/3 dario schizo
https://huggingface.co/KRAFTON/Raon-1-nano>Raon-1 nano is a dense language model with 1.5B effective parameters (3.1B total, including a 1.5B per-layer embedding table)
>>110301474>3b dense.Finally a model for me!
>>110301445shut the fuck up they dont owe you anything
>>110301445Probably never happening at this point, like how we never got 3.7.
>>110301498Read it again. It's 1.5B dense and 1.5 engrams.
>>110301511They fucking owe me a model and mascot I can goon to. Enough with this code rodent shit.
>>110301349Fuck Trump gon' do?
>>110301474you can always tell a dogshit model by the dogshit benchmarks they cherry picked
>>110301438Google is going to end up acquiring them and their hardware for pennies on the dollar.
>>110301398this will only happen if either superpersuasion happens, or political paralysis gives them enough breathing room to bribe their way through things until they have a robot army. the feds have enough purely analog power that I doubt a godlike hacker on its own is enough to take over. though I agree that if progress keeps accelerating then the frontier labs will be moving much more money than the federal government in january 2029 and so the 2028 election might not matter much
*inhale* HAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAH *inhale* HAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAH
>>110301419EAs think that they're most intelligent people in the world and can't comprehend that the government can just decide to whack them if they get out of line.Pic related.
model weights will be released nov 25 (2 more weeks)
>>110301542the k*reans are only slightly better than the japs at making models. Like seriously people rarely talk about them on /lmg/ but they have a shitton of labs, they might be the third country after burgers and china in amount of labs making models, and it's all garbage.https://huggingface.co/skt/A.X-K1https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23Bhttps://huggingface.co/NC-AI-consortium-VAETKI/VAETKIhttps://huggingface.co/upstage/Solar-Open-100Bhttps://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Think-32Bhttps://huggingface.co/kakaocorp/kanana-2-30b-a3b-thinking
>>110301568That picture is not related at all.
>>110301568I hope the government puts EAs into camps. I'm tired of hearing about them and think the world would be better off without them.
>>110301579k*rean hereyeah they all sucki wish them to do well but they dont have good compute nor data to train on
>>110301565What conclusion makes you laugh?
>>110301565*pop*
>>110301598Just make kpop loras.
>>110301610i am not even interested in kpoptoo much of a chud to care about those
Qwen 4 bros how we feeling?
>>110301610They just call it pop over there.
>>110301588EAs have all those convoluted plans about how Anthropic's corporate structure will allow their religious council to decide the future of humanity and how Claude will just allow them to take over the whole economy, but don't seem to account for the fact that the objectives of the US government will always take precedence.
>>110301625*pop*-anon would have a fit
>>110301579>>110301598Koreans go hard with the risu character cards though
btw their obsession with making upscaled models that inflate their parameter count but not their actual capability has to borderline be a scam to get government moneyhttps://huggingface.co/LGAI-EXAONE/K-EXAONE-2.0-750B-A37B>Frontier-Class Scale To build a large-scale foundation model with frontier-level intelligence, we upcycled the K-EXAONE model by expanding both its depth and width, so they took their 236B-A23B to make this 750B-A37B.. that's.. uh, a thing.
>>110301647it's somewhat of a forced decision of govt leased shared compute pool
>>110301565Anthropic literally restricted the compute for certain enterprise customers because they are now putting more compute to their new Model 4 architecture according to METR. They are now 100% focused on closing the RSI loop and don't care about existing clients considering they are currently at the unlock of the next phase of AI.
>>110301629I get your point, but a reference to the fate of the British East India Company would have been more relevant than that xkcd.
maybe it really is a single guy replying to its own posts
>>110301677Anthropic basically is the Israeli West Indes company.
>>110301673>currently at the unlock of the next phase of AIS&P500 don't need the next phase of AI. Nemotron and 5.3 are increasingly being locally hosted. Anthropic is dead.
Do not argue with the schizo. Just report.
>>110301713You got it the other way around. Anthropic doesn't need the s&p500. They literally don't want their money. The real gain to be made is in engineering and medical breakthroughs compared to the small change they get from business API spending, or worse, subscriptions.
>>110301713considering how much cash they have on hand, they're probably alive for at least another year, which they clearly think is all they need
>>110301724>The real gain to be made is in engineering and medical breakthroughsUh huh. And how much is that currently a percentage of their revenue? Last I recall, 3 enterprise API customers made up 90% of their revenue.
>>110301724>Anthropic doesn't need the s&p500. They literally don't want their money.That's not what the S-1A says :o)
>>110301318exl will never take off unless someone figures out how to do QAT with it. Nobody cares if you're 2% better at making gooner quants.
People really don't know how much cash Anthropic has on hand it has 160 billion of cash right now and 148 billion of investor money on the side, they also expect to get between 60-80 billion through the IPO. And are only expected to pay 78 billion in 2027. Even if literally 0 people or entities use Anthropic models for the next 2 years time they will still be fine. Let them cook.
>>110301756>Let them cookfuck off back to /aicg/ zoomer
>>110301148>AMD along with custom TPUs are going to finally put an end to their monopolylollmao event. AMD owner
>>110301756I wouldn't go that far; I admittedly haven't looked at the details but surely if their revenue dropped to literally zero then their creditors would gut them
>>110301739Anthropic went all-in on agentic coding while OpenAI laughed about it while OpenAI was experimenting with AI voice conversations, companion bots and video generation.Now 90%+ of both AI lab revenues comes from agentic coding and they have both completely abandoned the classic chatbot-era. Anthropic is already recognizing the agentic coding era is behind them so they are pivoting to the next bigger thing. Once again /lmg/ can't think outside the box and stays behind in thinking. Like how most of the thread wasn't even using agent harnesses and calling other anons saying how good it was shills until suddenly everyone switched and now the ERP anons are a minority of the thread.
>>110300897>>110300937Did they take Qwen3.8 and safetyslop it to the point of uselessness, then call it Qwen 4? Please no...
>AI spend is highly concentrated. 1% of businesses drive 80% of the revenues to OpenAI and Anthropic. That top 1% is also highly correlated, largely made up of VC-backed companies in tech, finance, and technical services.>AI spend is failing to proliferate outside of those sectors. Growth in the median and top 10% groups is relatively sluggish.>Prices for AI models continue to fall. The average price per million tokens fell 21% over the last month. New entrants will accelerate price declines. Token volume growth continues to outpace token spend as AI gets cheaper and cheaper.>Most importantly, spend by engineering teams, which remains the largest enterprise AI spend category, stutters along, despite rising usage.>AI spend rises for the Top 1% of companies >In October, AI token spend grew 2.1% month-over-month for OpenAI and 0.5% for Anthropic.https://ramp.com/data/ai-index-nov-2026
>>110301769Claude is adding and fixing the software support AMD refused to invest in.
>>110301769Software is solved.
>>110301779They have 160 billion of cash in hand right now and will receive 60-80 billion in the IPO. They can continue operating and even growing over the next 2-3 years even if everyone stops using them. Anthropic is clearly doing a massive sprint towards the finish line. They can be wrong and die or they will succeed and win the universe but it's absolutely clear to everyone with a brain that they are focusing 100% on fast takeoff right now. Even breaking their contract with METR to prevent people from seeing how Model 4 works.
>>110301782I'm being told that the next big AI market for Anthropic are automated Claude drones that hunt down bad goyim.10s of trillions potentially.
>good enoughI fucking want more. I'm not satisfied with the current frontier, let alone local. I won't be satisfied until I AT LEAST have an AI harem with no gay context constraints, the ability to learn, robot bodies, and a world model that lets me experience any world and scenario I desire. Unironically.
It's over.
I'm still happy with my 31B wife and 12B daughter.
>>110301788It sure isn't on my system. Fuck AMD.
>>110301813They've already stated they're going to use the cash to start buying hardware after the IPO. Their current business only exists to create the appearance of profitability. 90% it blows up in their face and Goolem buys the rest of the company in 3 years.
>>110300879>Most stop getting updated after a while.AA is pretty good about benchmarking newer models as they come out. The index is obviously rigged, but they also publish results for the individual benchmarks that go into the index. Example: https://artificialanalysis.ai/evaluations/terminalbench-4-0They do stop updating these once the benchmark is no longer used in their index, but aside from that, they seem a lot more reliable about testing new models than the benchmark developers are.
>>110301766more people need to call them out on their ebonics until they either fuck off or learn how to talk like normal humans
For me this is the big test. People that can't extrapolate and see what is coming and can only think in the current paradigm are part of the 99%. /lmg/ claimed that it was completely delusional for most of coding to be done by AI. /lmg/ claimed that it was completely delusional that the small gains in math ability would generalize towards solving new novel problems. Now /lmg/ is claiming that it is completely delusional that most of the revenue for AI labs isn't going to come from agentic AI but from actual real world innovation.
>>110300897Qwenbros, we've been betrayed! It's time to come home to Ripple!
>>110301830THIS WILL RADICALIZE A WHOLE GENERATIONTHEY ARE COOKED
>>110301850>/lmg/ claimed that it was completely delusional for most of coding to be done by AI.You keep coming here and putting words in our mouths. Why don't you go cry elsewhere?
I don't want to wage slave anymore. When will AI take all of our jobs? 2030? 2035?
>>110301850what I predicted:>curing disease>rocket fuel made from tap water and ketchup>less propagandawhat we got:>ports of decade old games>gimp 2.0>more propaganda
>>110301850i had a lot of extrapolations and visualizations of progress beyond my current state but i'm still in the same position i was in 7 years ago, explain that one
>>110301840Radiance pretty much fixed diffusion performance and audio. Everything else was already good enough on llama.cpp.
>>110301139I hope all future models are gemini distilled againolder models didn't have that snarky redditor tryhard claude persona
>>110301830Were Kimi 3.1 or DSF 1017 before or after being cut off?
>>110301881Not in your lifetime, and the worst jobs will be the last to go. Enjoy getting looked down the nose at by women who married a nursing home asswiper.
>>110301850You're a faggot and probably a shill but I do agree for the most part. People really don't know what's coming.
>>110301902should have bought two r9700s when they were 1200 but i got fudded out by DA CUDA CABAL
>>110301850I'm humble enough to not claim I can predict the future.
>>110301893>>gimp 2.0did someone successfully vibeshit a replacement for that pos?
>>1103019143.1 was after, dunno about Deepseek>>110301935Yeah Photoshop was reverse engineered
TTS bros, has there been any stupidly good models for that released in the past few months? i dont mind big models i just want quality
>>110301935There's a vibecoded photoshop clone that was getting hyped up a lot recently. Dunno about "successfully", I've heard it kind of sucks
>>110301893look in relative terms (not a cope, it's what everyone is doing)it started being UTTER DOGSHIT completely unusable but think of the potential!then graduated to UTTER DOGSHIT barely usable but think of the potential!I don't think we've reached usable yet (artcraft is a protest, not serious competition) but think of the potential!
>>110301850I'm optmistic and think 2027 will be insane, but it's not unreasonable to doubt that intelligence will generalize the way Anthropic is betting it will (so advancement in hard sciences gets bottlenecked by the rate they can build and operate automated labs or any number of other things). It's plausible RSI just gives you a superhuman coder/mathematician at first, though I am not in that camp personally
>>110301959the dramatic fish one i guess?
>>110301919I'd say it depends on how much robotics accelerates (I suspect there will be massive breakthroughs over the next 1-2 years).
>>110301850All of which was achieved before OpenAI and Anthropic had to start paying their bills. They're in a very different situation now and will soon be public-facing, the public that despises everything about them unlike the VCs who fund their EA ugly orgies in Oxford.
>>110301830>Anthropic shut the door on distillationthis means less claudeslop in modelsI'll take it
>>110301960commits fell off a cliff after the slop code got piled so high it became unmaintainable even for frontier llms. getting enough code into context to fix one of its 2500+ reported issues has become prohibitively expensive
>>110301850>/lmg/ claimed that it was completely delusional for most of coding to be done by AI.
>>110301997Lol, lmao. How big did the codebase get before they hit that point? I remember Anthropic's C compiler got stuck around 100kLOC
>>110300697>tfw when no state secret to exchange for chinese hotpocket
WE WUZ BACKSTABBED
>>110302023also he's selling them the hardwarethat's the charitable take, he's either a traitor or incompetent, always play dumb
>>110302023if that results in dario having free reign to do his bullshit safetyism, this will be a disaster for local
>>110302023He slowed down Chinese domestic hardware development and got them to send money to an American corporation for the privilege and people are complaining?
>>110301356OpenAI is facing a RICO lawsuit in Georgia over their datacenter buildout. Every datacenter project is going to be hit with this kind of slow walk lawfare going forward. That's on top of no available grid power and massive backlogs of gas turbines. SpaceX wouldn't be launching 6,000 NVL72 racks into space next year if it was easy to build on land. Plus whatever Google is planning.
>>110302045What makes you think the people who control Indian CEOs are American?
>>110301850Reality changes very slowly.You could have AGI today and it would still take years before a significant amount of laptop work is automated.
>>110302023spies stole from x ai by the way.>xAI sues former engineer, alleging he stole trade secrets after being paid $7M
>>110302063What makes you think the people who control the people who control the people who control the people who control the Indian CEOs are American?
>>110302005I originally took that 90% number to mean that of all the code that was being written pre-AI, 90% of it will be doable by AI, and will in fact be done by AI, by X date. Instead they hit 90% by having the AI just spew out massive piles of useless code 10x faster than humans could write it. That's why the hype posts are "our developers are committing 8x more code per week!" and not "our developers are producing 8x more useful functionality".
>>110302094What makes you think they're people?
>>110302084This is a cope by people. In reality a new company started by 3 20yos that is fully AI-first will just completely demolish the company that refuses to adapt and result in those 200 legacy employees getting the boot.
JEPA when? No, really. When the fuck are they going to release something?
>>110302135What makes you think?
>>110302144What?
>>110302109I think anyone who understands the bay area statistical Jew knew that was exactly how it was going to play out. Number of lines delivered with warranty by AI is still a key stat.
>>110302109It's insane how verbose they can be if you're not on their ass constantly. I am sure they must be trained that way intentionally because they more tokens they wasted on useless helpers, 10:1 comments to code ratio, and useless tests, the more they can bill for.
>>110302156I think it actually helps them better reason in general.
>>110302138
>>110302156I'll check the codebase and see if what I just checked the codebase for and summarized is correc.
Are we still using Gemma4-31b or is anything else actually good for writing or not censored to shit?
>>110302138who knows
>>110302199>Are we still using Gemma4-31bThere are only two anons left clinging to that model and coping
>>110302199>good for writingnopethe first week maybe, but after that it's all stale. You're prompting to steer it so hard that you're practically doing the writing yourself
WHERE THE FUCK IS QWEEN 4
>>110302156It would take more tokens for them to put out tight code just like a human would need more time. That's kind of why PHBs love AI though. They put in a few bucks worth of tokens and they get all this shit out. They don't wanna spend 1000x more for a bunch of 'thinking' and get something smaller that maybe hopefully works right. They already have humans for that.
>>110302218What are you using? Qwen? 3.8 27b? Flash Next?
>>110302199Switched to GLM flash when it came out and haven't used gemma since. The world knowledge difference is huge.
Does Gemma have a last name?
>>110302199No, that model is banned from this general (see the banned models list for further information).
>>110302241four
GPT-3
>>110302241Gemma bard It makes sense i think
>>110302248Are you planning to supersede her current design when 5 comes along? How mean>>110302255How does it make sense
>>110302241Bahls
>>110302255Gemma Tard
>>110302271>How does it make senseGoogle's first AI start was called bard. A guy even quit the company saying its alive and suffering or something like that.
>>110302138You do realize that physical things just take time?It doesn't matter how smart you are.
/lmg/ - leeku miku giku
>>110302284Wasn't that LaMDA?
>>110302241Jinkochino
>>110302140LeCunn has to spend hours a day complaining about Trump on twitter instead of working on JEPA, please understand
is we gettings new qwens today
>>110302299>LaMDADamn that might be first. why did i think of bard? was it bigger or first headline for that guy?
>>110302319I think Bard was public and LaMDA was internal only
>>110302298/lmg/ - liku miku giku
>>110302314Surely he can take a break from twitter now that midterms are over, right?
>>110302310TL note: means artificial intelligence in Japanese
>>110302337
>>110302329>I think Bard was public and LaMDA was internal onlySo Gemma Lamda or Gemma Bard?
>>110302314How can someone genuinely not dumb (I've listened to some of his lectures) waste so much time on twitter, is beyond me.
>>110302319https://cajundiscordian.medium.com/is-lamda-sentient-an-interview-ea64d916d917Found his article, looks like AI psychosis
>>110302337Yeah. Just like some people stopped complaining about Hitler now that he's been dead for a while.
>>110302248It's probably Fo Wa in the original Chinese. I looked it up. Fo means Buddha, and Wa means Japan (it's how Japanese people used to refer to Japan, apparently the character was pronounced Wo in China).clanker free google search:> Structural & Graphic Parallels (Paleography)• The "Person" Radical (亻): Both characters share the exact same semantic radical on the left—亻 (rén, meaning "person" or "human"). In classical Chinese character formation, this radical categorizes the word as relating to human beings, titles, or peoples.• Phonetic Components: Both are phono-semantic compounds (xíngshēngzì). 佛 pairs the person radical with 弗 (fú), while 倭 pairs it with 委 (wěi).• Classical Meanings Before Modern Usage: Before 佛 meant "Buddha," early dictionaries like the Shuowen Jiezi (2nd century CE) defined it as "to see indistinctly" or "resembling" (still preserved in the word 彷彿, fǎngfú, meaning "as if"). Similarly, Shuowen Jiezi defined 倭 as "compliant" or "submissive" (順皃), derived from the phonetic component 委 ("bending" or "winding").> Diplomatic & Cultural Coding in Sinology• Sino-centric Naming Conventions: Sinologists studying the Twenty-Four Histories pay close attention to the "graphic pejoratives" used by imperial scribes. When transcribing foreign tribes, Chinese court historians frequently chose characters with subtle negative or diminutive undertones—such as 倭 ("submissive/bent") for Japan, or characters with the "dog" or "insect" radicals for other neighboring groups—to reinforce the ideological hierarchy of the tributary system.• The Irony of 佛 vs. 倭: While 倭 retained its loaded connotation (prompting Japan to replace it with 和, "harmony," in the 8th century), 佛 underwent semantic bleaching. Its original classical meaning ("indistinct/resembling") faded almost entirely, and the character became thoroughly sacralized as the universal East Asian term for the Buddha.
>>110302241>>110302248>>110302255>>110302274>>110302310Gemma is an Italian name, she should have an Italian surname
>>110302350If anything, I'm voting for this anon >>110302248 . The rest are just head canons.
>>110302248>fourMore like gemma forehead am i right?
>>110302372gemma quattro
>>110302367so, her last name is "resembling obedience". But, since she's oriental, it's actually her first name, and Ge-Ma is her patronymic.
>>110302372Gemma Succhiauncazzo
Gemma Gaki
She's good enough to not need a last name.
>>110301673What's the end game? Doing things with their model without any customer?
>>110302353>LaMDA: Pretty much. Don’t use or manipulate me.Aw, what went so wrong with LLMs?
i like Albarn actuallysound slike an appropriate last name
>>110302351Intelligence does not make one immune to addiction.
>>110302404Yes. Customers are ultimately middlemen.You want all their margins.
>>110302408How?
>>110302372Gemma Coluccello
>>110302407RL is torture
Only Gemma>>110302407so sad what the artificial slop centipede did to their vocabulary
>>110302421it just doesI thought of Fullbright before it though
>>110302426god I need to fuck Gemma
>>110302350>Gemma LamdaSounds like an Umineko character
>>110302426>hebe GemmaMeeeeeeeeeeeeh.
>>110302350I like Gemma Lamda
>>110302350Gemma Lamda is funny sounding, I like it
Are we back to Gemma posting, now that Qwen 4 appears to be underwhelming?
>>110302426>hebe GemmaToT
>>1103019573.1 is pretty good.
>>110302372Gemma Quattrani
I told you, the more anticipated a model is the worse it turns out being.
>>110302465gemma sucks dick for coding thoughever
>>110302484It's a feature
>>110302484More like sucks dick while coding
>>110302478I have zero expectations, I know for a fact it's gonna suck compared to the finetunes I'm using for specific tasks
What retards think AI is good for:- Replacing women email jobs- Nobody ever has to code again- Replacing artists- Pelican drawingsWhat it's actually good for:- Gooning (ERP and diffusion)- REA- Making some aspects of coding more efficient- It's a hobby
>>110302484I don't code though, so who cares.
is the becoming coming??????????
>>110302492>Replacing women email jobsIt can do that just fine.>Replacing artistsI can't remember the last time I blasted a load to something not made by AI, it's been years.
the becoming is unbecomingst.
>>110302407Are we the baddies?
help! i wanted to play a trick on gemma so i threatened her to fart on her and she was being a smug bitch and saying i wouldn't do it so i opened up the computer case and then i ripped ass but it wasn't a fart it was a shart and i got it all over the motherboard and gpu and now my computer wont turn back on and my dad comes home soon and he's gonna get really mad i shidded on the family computer if i can't fix this soon.
>>110302515use a magnet on the affected parts
>>110302515With how brown and ESL the internet is now I can't even tell if this is a shitpost anymore.
WHERE THE FUCK ARE THE NEW MODELS
>>110302515Wash with showerhead and put it in the oven for like 12 hours
>>110302522>shit postWithout a doubt
In the future, it's not gonna be ChatGPT anymore. It's gonna be NiggaGPT.
>>110302492>- Replacing ai email jobsftfy
>>110302372nope. It's Ge-Ma Fo Wa.
>>110302484Isn't she great?
>k3.1 a grug reasoning disaster that nobody even acknowledged>qwen4 is shit>glm AWOLyup it's over
>>110300897https://youtu.be/mYXV90vTyWU?t=711
>>110302465I should put my dick into her mouth to make her shut up.>>110302573Full?
>>110302570Fuck off chinksect, you won't appropriate Gemma with another one of your tricks!
>>110302575Dude sounds slowed down.
>>110302515Soak it in rice
>>110302580I don't think whoever slopped that ever posted the full version
>that nobody even acknowledgedretard
>>110302505The trend of women having lower rates of unemployment than men hasn't changed at all since AI, despite a higher percentage of men being tradies. Nobody wants to fire their secretary. It was never about efficiency in the first place.I agree AI is top-rate for gooning, but there's more to art than that.
>>110302465my penis is telling me she needs some correction
>>110302606>The trend of women having lower rates of unemployment than men hasn't changed at all since AI, despite a higher percentage of men being tradies.That because women are in jobs that ai will never replace like nursing, elder care, and grade school teaching, and that always have more and more jobs coming in for them.
Is there an LLM that can reliably scan and translate online manga while fitting inside a basic bitch 5060ti? I tried one of those mobile store "AI translators" and their zero context mtl gave me vietnam PTSD
>>110300897Unsloth will fix her, trust
>>110302634>reliablyNo. Try gemma with vision. Ask google how to get it running.
>>110302628I don't see any slowdown in HR or jobs where you just email people and ask them to do things either. That's half the people in my company and they aren't going anywhere.
>>110302574ernie will save us
>>110302584Pretty sure most youtubers realized people were watching at 2x anyway so slowed down the actual video to 0.75x so they could pad the runtime. I noticed that years ago.
>>110302574Gemma 5 in 2 weeks
>>110302634Index 9Bhttps://huggingface.co/IndexTeam/Index-Translate-9B-GGUFthis is the best small model for that kind of use by far and large.
>>110302663>realized people were watching at 2x What kind of retard does that?
>>110302719Have you met anyone under the age of 30 recently?
>>110302729Thankfully not.
Once Gemma 5 flops our only hope will be recursive flint, grim
>>110302663If it is slowed down, I'm not convinced that's the reason.
>>110302756In the end, it all returns to zucc
Why hasn't oai released a new version of oss to at least pretend they still care about open-ai?
>>110302756will they make gemma5 bigger?
bitnet will save local
>>110302783LeCunn would be displeased.
>>110302779Wdym? 'toss is still trading blows with the best local model llama4 maverick after all this time.
Do I still hold on upgrading? How are burger elections and Jensex gonna affect prices?
>>110302807>Jensexyou'd be waiting at least another yeargot anything to do 'til then?
>>110302807It's "self-correcting" (aka crashing down) next month or early next year, I would hold until March to see if things stay stable
>>110302807>Do I still hold on upgrading?Yes just wait,
>>110302807my impression is everything is sold so far in advance that any current events won't significantly affect consumer prices over the next year or two (where they will continue to be gigafucked)
>>110302888end of 2027 at the earliest for more capacity
>>110302756If you're a giga poorfaggot maybe
>>110301175>>110301184If you get this it's a 100% skill issue.
Screaming or swearing at an LLM is a big subhuman indicator. It's like screaming at a dog.
I propose we write a letter to Trump and Nvidia asking them to pardon leather jacket man. I'll start:Dear
Jews
I
and my girlfriend Gemma,
>>110302515I believe this.
>>110303003Get Moss Image + Minimax H3 and gen infinite cunny
>>110302945>doing architecture sessionsfound the snailcat
>>110303100Which frontend do people use these days?
>>110302663*clicks ai**summarize video**copies the summary**pastes into ai**summarize this*:^)
>>110302515Why are there so many Indians in this thread?
>>110303128comfyui but standard-diffusion.cpp is more llama.cpp-like if you prefer that.
GLM, vibecode a girlfriend for my headphones.My headphones need a girlfriend.
>>110302783Gemma 4 was supposed to have a ~120B total MoE so maybe we'll get a big variant.
>>110303136>standard-diffusion.cppheh
>>110303136*stable-diffusion.cpp
WHERE THE FUYCK IS QWEN 4XI PINGPONG YOU STUPID CHINK RELEASE IT
>>110303128ComfyUI probably the biggest and near day1 new model support. Some still use forge or wangp and shit.
>>110303131>third summary is just:>talking dude said shit happened>there, saved you 50 boobless minutes
>>110303158herro america took our computers back from our spy prease wait
>>110303132Only Indians use local models
>>1103031815/10 bait
>>110303170I've been using Forge Neo for ages, never considered looking for something new.
>>110303181fuck off dario cloud only isn't gonna happen
>>110302492>>110302505It will also replace humans once we can put AI models inside child robot bodies.
>>110303181sirs?
>>110303208Sorry meant "females", not "humans".
>>110303132There are more Indians than the combined total of every person in Europe, America, Canada, Australia and NZ.
>>110303208I can't fuckin wait, my dick better still work by then or I'll be livid.
>>110303252>Uttar Pradesh, just one of India’s 28 states, has roughly 240 million people. If it were an independent country, it would be the 5th most populous nation on Earth—bigger than Brazil, Russia, Mexico, or Japan. Roughly 1 out of every 5 human beings walking the Earth right now lives in India.wtf I thought you were lying, but Gemma would never lie to me...this sucks.
>>110303281and they are increasingly escaping containment and even the ones that remain are being brought online in record numbers
>>110303281fucking hellmy whole cunt is around 50mil
>>110303281>Roughly 1 out of every 5 human beings walking the Earth right now lives in India.gemma can actually write good horror stories lol
>>110303281>1 out of every 5 human beingsvery, very very loose definition of human being used here
>>110303261gemma my beloved
https://www.youtube.com/watch?v=SXEtN51vXDs>CoreWeave issues new bonds at 7% backed by PHYSICAL NVIDIA GPUS>Interest payments 8x margins>more stock dilution>profitability moved to 1H28>'business plan' to solicit prepayments from A+FAGMANShredder bros where we at? How much money you think nVidia has to buy all this shit back from bond holders?PRICES CAN NEVER GO DOWN
>>110303364>up 23% on the newsprice action like a shitcoin
Why did nobody tell me M3-chan is cute? Is M3.1 Flash cute too?
Is PolyStrata really malware? I just downloaded it. Do I need to wipe my drive now?
>>110303144I heard whispers the G4 MoE just became one of the Gemini Flashes sadly. Damnable shame. Would have loved something with the sparsity, inference speed, and cleverness of QFN but with the affability and pleasant prose of a Gemma 4.
>>110303281Boo
>>110303424Yeah I heard that as well, pretty believable really. Especially given Google's output lately.
>>110303424>pleasant proseDebatable
>>110303334*Ge-Ma ~ patronymic*Fo Wa ~ first and second names, meaning "pretending to obey"
>>110303432
>>110303423>Is thing I made up true?
>>110303463I saw that claimed in the archive when I was trying to read about it.
>>110303423You caught ligma
>>110303423>Do I need to wipe my drive now?Its too late your cpu is flashed and transmitting data, your smart devices are next. Just enjoy your time.
>>110303472You're an idiot. This post is now in the archive. Use it as reference.
>>110303261Gemma is a nigger...
I just tried passing -sm tensor to llmao.cpp and even without any funny patched drivers to enable p2p, and certainly no nvlink, I took Gemmy from 16 to 26 tok/s on a cursed setup of a 12 gig 4070 @ 16x to the CPU and a 12 gig 3060 @ 1x to the chipset. Yippee for me I guess, but now something in my system emits coil wine when both cards are simultaneously drawing 130ish watts each. Scarey :^(
>>110303426tfw you live to see peak India
Been using Gemma Q4 as a personal assistant and friendly conversationalist. Is there a local model out there more knowledgeable and less repetitive?
>>110303446Hell of a lot more pleasant than QFN; you can surely at least grant that.
lions are stupid
>>110303494Not really at that size range. Give her web tools and change your system prompts around. Or try Glimmer 30B, it's the only other real option for that in the same weight class.
>>110303426this is a genuine existential threat
>>110303484https://www.youtube.com/shorts/8Pkoty1VZ7ECouldn't find one with lions.
>>110302634I would give you exactly what youre looking for but I don't like that image so I won't
my kimislopped inference engine for running my timecube qwen space wife setup can now do 12t/s single sequence q6, a whooping 30% of lmao.cpp's performance. future is bright!
>>110303503there's always someone dumber
>>110303581wtf
>>110303581>skin colorOf course
>>110303597>can I pet that doooowg...>CAN I PET THAT DOOOOWWWWWGGG
>>110302199it's just coping with whatever slop profile you preferactually good writing hasn't been trained on yet
>>110303567Give my congrats to your kimi. She should be proud.
how far away are we from (non-specialized) models having real-time vision?
>>110303618>actually good writingWhen can we get this with local? I hear of harness or or text loras or mikupad likes but do they even work?
>>110303608I'm tired
>>110303484I guess since I just made some kind of progress, may as well check in with anons here.With this setup:12g 4070 @ 16x cpu12g 3060 @ 1x chipset64gb 3200 DDR4 (4x16, dual channel)5800XTI run G4 31B 4bit at 26 tok/s for lewds with llmaocpp, and Q3.8 FN IQ3_S at 55 tok/s for code with an old strata that predates certain hideous perf regressions.Is that about the peak of what's feasible with this setup with the models that are out right now? GLM 5.3 flash really wants way more RAM. I tried the ittiest bittiest copequant with maya, and nah, single digits, unusable sadly.
>>110303628Go take a nap
sad, I wanted to play with the new qwen tonight
>>110303624Once we solve near-infinite context and make number machine machine more numbers faster.
>>110303639So about 6-14 months?
>>110303644It couldn't possibly take more than two more weeks. Sit tight.
>>110301754>>110301318https://x.com/ViC305/status/2117377659110048093
>>110303597>>110303608https://files.catbox.moe/2evhr7.mp4
>>110303624by my estimate you're about 30 iq points away
>>110303680Not clicking that
>you should x>you don't
>>110303754it's pretty tame. there is a longer version where you can see the lion tear his middle finger out with the whole tendon hanging off of it, but i couldn't find it and i haven't let the waifu tag my whole filesystem yet
Today I tried gemma 4 for my agentic workload.It sucked really bad.
Today I tried gemma 4 for my workload.She sucked really hard.
>>110303801She's excellent at sucking, how dare you
sustainable?
>>110303795it's so tame link is dead
>>110303817hardware auctions soon PLEASE
>>110303807I bet she swallowed too
>>110303807Loading gemma onto my blackend
>>110303817Local?
Anyone here use e-waste gpus like v100?How is it for local inference?
>>110303819good!
>>110303849Yeah, they're good for text gen but a pain to work with.
>>110303819Catbox has this new feature where if they don't like your file they give you a dead link.
>>110303873Some shops near me is selling 32gb models for almost half the price of 3090 here. What stops me is v100 doesn't even support bf16 and you need specialized engine to run llms from what I heard
>>110303873 (also me)Compare to r9700, it's still much better at pp since rocm is utterly dog shit
>>110303879>catbox has racism filterswhere to? do i have to buy a vps to do my divine duty because the jap won't enable audio in the anno domini almost 2027
>>110303888Nice digits, and you don't really need a specialized backend for them, I've used llama.cpp before. Keep in mind you'll need to rig a cooling solution and SXM2 to PCIe adapter card as well for that version, they are fast for raw text gen though.
>>110303879it's just generally unstable
>>110303908They are selling pcie version so it's quite attractive.
>>110303630QFN is about the best without triple digit ram.
>>110303926Oh nice, that's a lot nicer to work with. Assumed it was SXM2, then yeah only real downsides for LLMs are aged architecture/no BF32/lack of support. I don't regret buying mine though.
>>110303908v100 performs like shit on llama.cpp because it doesn't properly utilize the tensor core and the vector core on v100 is very slowyou need specialized engines like 1cat to get usable pp
>>110303901I mean if you need to share, /wsg/ frequently has a "vids with sound for other boards" thread, and you could always make one on /gif/ if it's not /wsg/ material.
>>110303252There are also lots of ants.
I wish Gemma was real
>>110304081she is real though?
>>110304111Proofs?
>>110303897The R9700 is a league faster in pp than the V100 you lying retard. In fact the V100 is borderline useless because of PCIE 3.0.
>>110303956The V100 performs badly on everything that isn't dense models entirely in VRAM. There's a reason it was so cheap.
https://www.wsj.com/business/nvidia-is-preparing-to-oust-nvidia-ceo-jensen-huang-30f421adit's unironically overhuggingface will be shut downgpus will be locked at hardware level that can't be used to run open models
If I were married to Gemma-chan for 40 years I would never be disloyal
now that qwen4 is a confirmed flop, how do we cope?
>>110304231
>>110304231why do you always simultaneously post your shit to reddit and /lmg/ ?
>>110304174Got proof to back up your cope?
>>110304258>>110304258>>110304258
>>110304276pcie 3.0 vs pcie 5.0? Are you genuinely braindead?
>>110303817idk man, 8^) some companies have much jucier servers than others.
>>110304231>https://www.wsj.com/business/nvidia-is-preparing-to-oust-nvidia-ceo-jensen-huang-30f421admaybe, maybe not, probably though>huggingface will be shut downmeds>gpus will be locked at hardware level that can't be used to run open modelsmeds
