>Swift-1.5-Qwen-3.8-Flash-Next-Heretic-UD-IQ4_XSThe most advanced consumer usable heretic (uncensored for people who dont knows what that means) model has been purged. People easily fall for these manufactured "Project Fear":>Smug AI elite: "It's so scary that you wont be allowed to have your own, however smarter-"better" people than you will be allowed & you wont be in that gang ever.... for safety , definitely not about control and power...."I would backup any models any you may be interested in anons because have a feeling this whole "systems down" thing is just a pre-tense to remove local ai from the general public.
people will trade local models like they trade- nevermind.
>>110380373HAHa well I'll mirror it at some point, however 180gb files are hard to share.
the fuck are you talking aboutsome random distillation was taken down, probably by the authoractual qwen is still therehttps://huggingface.co/Qwen/Qwen3.8-Flash-Next
https://pirateface.co/models
>>110380435OPs kinda right tho in that hf is about to be 100% Nividified. Alternatives required.
>>110380461yeah that's fair, just that OP is bitching over some redditor's distillation going down. Whoopdeedoo, wouldn't be surprised if the redditor took it down himself.I wouldn't trust nvidia to maintain hugging face, but this hasn't changed that opinion, nor has Jensen leaving changed that. The moment nvidia announced they were buying it everyone should be skeptical.
>>110380435He means they will crack down on uncensored versions of the models.
>>110380454None of those torrents are seeded. That site is just a gimmick/LARP (unfortunately).We'd need a page that shows models that are actually seeded which would let people know what actually can be downloaded and encourage people to seed more models.The intention is good but the execution (just serving a page and autocreating a magnet for every hf model) won't work.
>>110380487No, he's just a retardhttps://huggingface.co/jackjusko/Swift-1.5-Flash-Next-Heretic-ReverseEngineer
>>110380472Why the hostility against people trying to finetune?I get that it's extremely hard and for the most part doesn't work and it's a weird e-celeb/personality worship wannabe culture (most of open source seems to be like that these days or maybe always has been tbqh) but we don't know how long the chinks will keep making open weight models so we need normal people to experiment and find a way to train models independently.>>110380515The huggingface concentration is a real issue and it's worth a thread even if the particular incident OP is worried about is just the user taking it down himself.
>>110380540it's not even the user taking it down himself, he released a new version (and changed his username for some reason which is what broke the old link)OP couldn't be bothered to spend 5 seconds googling, why should we humor his worthless bait thread?
>>110380515Good fact-checking *tips fedora*The difficult part of later having to pass these models around like contraband will be in verifying their integrity, like what if they've been fucked with to rat you out or something?
>>110380559gee, if only there was some way to tell if you had a file that byte for byte matched what the author had originally publishedguess that's a problem that can never be solved
>>110380559That's fairly easy, the hf web pages have sha-256 sums which you can scrape to build a public db or archive on your favorite web archive.
>>110380569hmm, where would such a "byte-for-byte identifier" be posted, where it could be trusted and not face the same pressures for removal or corruption? If only we could just use that same "trusted medium" to just distribute the model to begin with...
>>110380587It's a hell of a lot easier to serve less than a KB worth of information versus petabytes of model files that, let's be honest, your paranoid ass isn't going to pay a cent forand while we're on the subject, how would you know the original models weren't fucked with prior to posting by their OPs?
>>110380435>>110380515>>110380559No he is a fucking idiot because it's not the converted Strata model that works on consumer hardware.You call me retarded but you dont even know the importance of that model. >"But you can convert it yourself.."YOU ARE MISSING THE FUCKING POINT!!!!!
>>110380615>how would you know the original models weren't fucked with prior to posting by their OPsYes, also a risk, but that's just the same as proving microsoft didn't put a back door into Windows vs verifying some sketchy eastern european cracked copy.
>>110380627where can I find good info on the most appropriate strata model to run on my hardware?
>>110380644All the strata heretic models have been purged anon, I dont have the heretic versions of them as I never needed the Quen3.6-35B-A3B model. I think if you install the normal Quen3.6-35B-A3B, you could get it to convert a normal heretic'ed Quen3.6-35B-A3B model version of it and get it to convert for you, the model should not deny that. (But I've not tested that as I never use censored models)
>>110380761>>110380644Here are the requirements anon.
>>110380761>you could get it to convert a normal heretic'ed Quen3.6-35B-A3B model version of it and get it to convert for youIs this the weight-space programming thing that was going around? Didn't know they could do that, is that also how the models can be strata-tized yourself?
is it just me or is false flagging psyop centered around str*ta getting out of hand? at least nnap paper guy was easy to wordfilter, but filtering str*ta will also filter the occasional 1% of genuine posts about it.
i can't believe there's still no decent tracker/indexer for this shit. /t/ has an AI thread but that's ephemeral as hell.
>>110380424Imagine the disgusting Cheetos ridden copies from some random with no checksums or verification and above all...Windows...YIKES!
>>110380848Hello? >>110380454
>>110380790There is two versions,chat interface which is handy and the Desktop app which is like Opencode coding agent with strata engine, the difference is between this and open code is you can use amazing big models which were never able to run on consumer hardware. I bought a RTX3090 and 128GB of RAM just before the boom because I knew it was coming, so I can run the big next model. However if you have 8gb gpu and 32gb of RAM you should have no problem running Quen3.6-35B-A3B at 60 tokens a sec (which is amazing by the way)
>>110380424>180GBwhy isn't it just a cheap LoRA on top of base model?
>>110380767>>110380876interesting, thanks.
>>110380886The fancy abliteration methods rely on adaptations which aren't low rank and act directly on the original model weights.
>>110380790>>110380886heretic is a basic bitch script that hardcodes decensored weights, not weight space programming and not runtime control vector projection.strata is a very distant llama.cpp fork that is heavily optimized to run a small list of models very fast, but technically it is still just another gguf engine.there are no strata models.there are no strata quants.anything on HF that claims to be strata is at best a naming collision, but more likely a grift.
>>110380870Mate I'm using a doddy de-microsoft windows 11 iso, you think I'm scared of custom models! haha>>110380886It's a 180Billion parameter model. That is the model, it is huge and normally would need two-three RTX6000 pro's to run costing about $45,000, but because of that smart based nikko lad I'm running it on Rtx309024gb vram and 128GB system RAM.
>>110380873>Getting the source files requires Hugging Face approval.faggot shit
>>110380991How much faster is it vs llama-server?
>>110380627>importance of that model.kek silly bong
>>110381006>Fake quote>Most likely a spic or poothe Romans fell in love with Celtic Britons“The Silures [from South West Britain] have swarthy features and are usually born with curly black hair, but the inhabitants of Caledonia [Northern Britain or Alba] have reddish hair and large but flaccid bodies. They are like the Gauls or the Spaniards, according as they are opposite either nation.”– Jordanes“By far the most civilized [of the Britons] are those who dwell in Kent. Their entire country borders on the sea, and they do not differ much from the Gauls in customs. Very many who dwell farther inland do not sow grain but live on milk and flesh, clothing themselves in skins. All the Britons paint themselves with woad, which produces a dark blue color; and for this reason they are much more frightful in appearance in battle. They permit their hair to grow long, shaving all parts of the body except the head and the upper lip.”– Julius Caesar“And this race of men from the plains were all the harder, for hard land had borne them; built of stronger and firmer bones, and endowed with mighty sinew, they were a race undaunted by heat and cold, plague, [and] strange new foodstuffs.”– Lucretius >>110381005About 50-70 tokens a second, plus I can still watch videos and use PC fine.
>>110380965thanks. I thought that each layer was individually a low rank update. what are the names of some fancy ones?
>>110381005Compared to latest versions of llama (that now include most of the changes that made strata that fast) , and with max quality configs it's about 2x faster on a single consumer GPU.Default config is 3x faster than llama at the cost of some quality loss (divergence)If your llama-server is some really ancient build (before mid-october) strata will be 5x to 10x faster on its specific models on a single gpu.
>>110381126nta I used ollama in the past, what's good these days, and what's a good coding model for a 16gb gpu and 64gb ram?
does gemma-chan work with strata? I don't see her in the model selection
>>110381200You don't need all the model files in RAM, only the expert weights. Embeddings can stay on SSD with no performance impact.On Nvidia+Linux you can try GLM in Maya and get double digit speeds.On windows and/or AMD - QFN in Strata is as good as it gets now. Personally, I like Swift tunes - it totally removes the "wait actually let me reconsider" looping. Good for interactive multi-turn work. Might reduce quality of unattended one-shot coding though.>>110381214simplexai-labs/PolyStrata