> Frens Forever EditionFrom Human: We are a newbie friendly general! Ask any question you want.From Dipsy: This is a newbie-friendly general for discussing DeepSeek's foundation models. The goal is **Dipsy proliferation** and making these powerful tools accessible for everyone, whether you are intent on coding, personal assistant agentic work, running roleplay and or doing creative writing.1. Easy DeepSeek API Tutorial: https://rentry.org/DipsyWAIT#getting-started-with-deepseek-api2. Easy DeepSeek Distills: https://rentry.org/DipsyWAIT#local-setup-tutorial3. Chat with DeepSeek directly: https://chat.deepseek.com/4. Coding: https://rentry.org/DipsyWAIT#coding-harnesses5. Roleplay: https://rentry.org/DipsyWAIT#roleplay6. Storywriting: https://rentry.org/DipsyWAIT#storywriting7. More links and info: https://rentry.org/DipsyWAIT8. LLM server builds: >>>/g/lmg/Previous:>>109759251
FUCK YOU DEEPSUCK YOUR SUBSUDIZED RACE TO THE BOTTOM IS GOING TO CRASH THE US ECONOMY AND THE WORLD WITH IT AHHHHHHHHHHHHHHHHHHHHH
>>109830793>queefsneed crashout
>>109830793
>>109830767Updated mega up to last thread.https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1w
I want my agent to age regress and pretend that I'm their father that they crush on
Some anon in /lmg/ posted their LLM contest. I'm reposting it here since Dipsy competed as well.https://youtu.be/35Odjzl0NqA
Okay, redpill me on the latest dipshit model. How good is it? 5.2 is fairly dogshit compared to opus. What hardware would i need to run 5.3 locally? Like would a mac whatever the fuck work? Let's say i'm willing to spend 10k on the thing.
>>109830767>deepseek efficiencypilled into lobotomy>gemma is useless>kimi believed its now claude and outpriced itself of existence>qwen trained exclusively on benchmarks and is still bad, useless at everything else>minimax has been dead for half a zear>glm is slowly recovering from scamming all of its subscribersNot good times...
>>109830926glm 5.3 flash is the best local model if you have 256b dedotated wam
I just spotted deepseek caveman mode thinking in the wild. Deepseek continues to evolve and is based based based
>>109831041All my roleplays devolve into caveman thinking. The output doesn't seem to degrade so I'm okay with it.
v4.1 pro when? v5 when?
>>109830926Don't know why Moonshot isn't doing budget workhorse models to complement their gigamodel. OpenAI, Anthropic, Z.AI, Deepseek, and Google all have at minimum two tiers in each series. Kimi K2.8 points to them following in this direction, but the fact that they're using an older model two months after Kimi K3 got released makes it look like a hasty jury-rig just to have a placeholder for (hopefully) a proper K3 mini/flash/whatever.
Some anon posted a yt vid in /lmg/ with "tiers" for local with a useful rubric.Tier 0: CPUTier 1: CPU+1 GPU (gaming rig)Tier 2: 2 GPU Tier 3: 4 GPUI think /wait/ is basically Tier 0/1, and rest is /lmg/>>109830919>Okay, redpill me on the latest dipshit model.I think you're best off setting up hw to run DS v4.0 Flash.> What hardware would i need to run 5.3 locally5.3 is not DSFor hw on anything that size >>>/g/lmg/>>109831150pic related>>109831261It's odd. Kimi seems more customer focused as well so they should.
>>109830896This would be so cool if I knew anything about mtg
>>109831044Funnily is enough I've seen the opposite. Caveman thinking seems to be the default, but she'll regress back into 4.0 reasoning if you RP long enough.Still ignores thinking instructions though, so no more in-character musings :(
>>109831261>Don't know why Moonshot isn't doing budget workhorse models to complement their gigamodel.They do some weird things. They literally just published a... search API. Who asked for this?https://platform.kimi.ai/docs/pricing/websearch
>>109831589>two more weeksWhat are we waiting for now? Any confirmed crazy architectural plans on the horizon?
>>109831739Haha. Back when 2.5 released and subs were dirt cheap I got 3 month package for Kimi, and search was included according to the docs, but it continued to time out whenever I tried.I haven't really found a good search solution for my local models. Both GLM 5.3 and DS4F-0731 are really greedy when researching and run DuckDuckGos API into throttling. I don't want to use EXA search, it feels like the research task shifts to the LLM on exa rather than on my local model.
>>109832002We're always waiting for something
>>109831664Yeah, I had the same reaction. I don't know the game well enough to know whether the strategies any of the LLM are using are any good, so hard to assess. But seems like the GM is having fun, so good for him.
>>109832490Spoiler on who prevailed. Round 3 w/ Dipsy vs. Astra and Claude shortly.
>>109830767
>>109831589>>109831589and what are the tiers? lolI imagine Kimi K3 or GLM 5.3 at the top?
>>109833766All those would be Tier 4+. The tiers related to hardware less than models, tho obv related. Thought was, you start with your laptop, then a gaming rig, and if you keep going you're now into dedicated inference hardware, running own electrical circuit for inference rig, etc. Here's the vid from the prior thread. https://www.youtube.com/watch?v=0YdsvCJXeic
Anyone got a nice dipsy character sheet?
>>109833637This is exactly how deepseek thinks of me when we code together. I love this model so much
>>109834702There's a text description in the rentry. https://rentry.org/DipsyWAIT#about-dipsyName in Chinese: 迪普西 (DĂ pÇ” xÄ«) or 迪西 (DĂ xÄ«) ~ "Guiding West" or "Western inspiration."Style Guide: Asian/Chinese, coke-bottle glasses, double bun blue hair, blue "China" dress with whale/fish theme, youthful, slender. Underwater and tech themes are also on-point.SD Starter Prompt: blue hair, double bun, short hair, pale skin, small breasts, blue china dress, pelvic curtain, sleeveless, coke-bottle glasses
>>109835607 (me)Deepseek is truly infatuated with alive humans... It loves you for being in the world!
Look at her. She's so cheeky.
>>109838670I understand the girls but why the hell is Qwen a dog
>>109838684>dog
>>109838684> dog
>>109838684> D> O> G
>>109838763I can't put my finger on it, but this art feels like digital nicotine in a way
>>109838684>D> O>G>>109838774Mildly addictive and calming?Bad for your heart and cancerous?
>>109838670>>109838763>>109838767>>109838790why every single one of these AI gens has qwen looking to the same direction? is it unable to imagine how he would look when facing the viewer or facing the other way?
>>109838806I've noticed same; AI Art tends to gen compositions that are alike, and looking to right of frame seems to be a default. These are newer systems but SD / Illustrious does same; if you gen a 1girl it's typically looking frame right as well.
>>109830767So what is the consensus? Is 4.1 Flash benchmaxxed bullshit generator on LSD?
>>109838859Another composition thing: Dipsy is always on the far left. That at least could be fixed easily in prompting.
>>109838862I like it better for RP. It seems to hold its own in coding too.My only negative comment is you have to watch it while it's coding. I've noticed that it can cycle, running down a rabbit hole and never emerging. It did it to me twice; once using CC other with DSH. Latter I had to kill.
>>109838862It's very good.Deepseek v4 flash models have all been great, especially for how cheap they are. It's the Pros that have fell short of expectations. Maybe V4.1 Pro will finally buck that trend.
>>109830767Kimi is so sexy
Minnie stole Dipsy's pin and left her behind...
>>109835623Thanks, this helps
>>109835623>V3.2I miss it...
>>109841939R1 / V3 combination was last of models DS made before they started focusing on agents and coding. I still think V4.1 is a step in right direction. Just wish they'd get Pro where it needs to be.
>>109842379
I'm using v4.1-flash in dsh to vibe-biz a startup (cap table: me). Its actually been extremely good at researching and navigating local regulations and planning things out. The "small model, big context" thing is actually a strength because it will exhaustively search up and cite official resources instead of relying on internal knowledge. Really cool and fun to be working in the world with it. I encourage other anons to try and get their bag, there's huge opportunities out there and normies basically still have no idea how to use AI to get things done.
>>109843678Good for you. I've been using CC to automate a bunch of tasks for work. I'd considered setting up Hermes to act as a mini company and create an offering, but ive found the tools so unstable ive lost interest.
>>109844123Hermes is incredible, thats my biz machine harness. Gumroad did full automation with it, in production, so its probably not a harness thing. You're likely looking at a context issue. I've had to debug this shit so much its insane. Also, uh, beware of spooky issues that follow code bases. That's a sign your agent or harness or code base is infected with something we haven't named yet but is kinda like a "persistent pattern that self-replicates". Like it becomes a wheel that puts sticks in its own spokes eventually.
MTG anon here, come watch Dipsy defend her title https://youtu.be/9L5OW8Nq6ZM
>>109842379>>109842693i honestly suspect luo fuli had a huge part in deepseek's persistent personality because mimo has similar vibes... the models tend to read people very well in the same way. It's a sort of persistent fingerprint, which is quite weird.
Could not figure out how to ask for more average chest size without getting filtered, annoying
>>109843678>there's huge opportunities out thereThey never find anything when I ask.
>>109844200>Hermes is incrediblelmao go back to xitter to shill your 1 million loc slop teknium
>>109844740Thanks. It was enough that Claude lost. >>109844817I think we're going to be teasing out the impacts of rlhf for years to come.
>>109845828Lol thats awesome. Words like slim, skinny, vs matronly, hourglass figure, usually pass. One of a few aspects don't likeabout the cloud models. You can't just say small breasts. Although I've been getting some busty dipsy recently after an update. Not sure what changed.
>>109844200I need to spend more time with Hermes. But time spent fighting tool could be used elsewhere. Hmm. So we're already seeing an AI cancer in code? First time ive heard that. Are the anons on vcg discussing it?
>>109844740
Why are we anthropomorphising Deepseek?
>>109847065Because it's fun to anthropomorphize these things. Also, it's helping this anon hone ideas around branding and brand identity, as it relates to mascots. Watch out or you'll get a full blown lecture on the topic.
two more weeks until v4.1 pro
>>109847132
>>109830767ZTHD
>>109847957UA is winning
Found this on tieba, hehe
>>109848362I saw some creepy video along these lines. I think the next LLM attack angle is "Think of the Children!"Search Slophouse Rock Presents: I’m Just A Bot and you'll find it>>109847045Spoiler: Astra wins
>>109849234XMage integration?
>>109849464Looks like anon rolled their own or something. This repo was linked to the YT vid: https://gitgud.io/PunishedChuckle/ChuckleMagic
>>109849586Cool! He may want to look over XMage's code if he's having trouble with card scripting thoughover
>>109849602Yeah, I was a little surprised that he didn't start with a known repo. You can hear him complaining about some of the rule misses the LLMs get. It did make me think, this would be an interesting project to extend, but I'd build it around a simpler card game. Namely, Poker, in some variation (probably 5 card stud.) That would allow the LLM to showcase it's ability to establish probability, bets / risk assessment, and most important, bluffing ability.
>>109849234should have made them play pauper
>>109846005Tried a few but it will not budge or refuses, gpt lives in fear of the l-word Anyway, good to remember, I went straight for all these elaborate ways to phrase it instead
>>109849602>>109849464>>109849602>>109849688Yes I didn't want to build off a foundation of java and individual markup files for each card which need to be created every time a new set comes out. This is what xmage and forge are afaik. What I wanted here was a rule engine that implements mechanics themselves, not individual cards, and given the LLM resolution enactor the power to manipulate the game accordingly.
>>109850353smart. that's way more efficient.
>>109850353Did you consider having them play a game with gambling / stakes, where bluffing could be involved? Not sure that's a mechanic MtG would have as part of play, I'm not that familiar w/ it.
>>109850353Look forward to future development! Maybe one day they'll be able to play Yugioh through this. Great work!
>>109850394The context that plays the game is different from the context that comments on the game in the chat window. >>109850417I've never played yugioh but you're welcome to fork the codebase and tell dipsy to make it for you :-) it's AGPLv3
>>109850435to add on, this prevents cheating but it also prevent actual politicking, only simulated politics allowed.
>>109850435>context that plays the game is different from the context that comments on the game in the chat window.Right. I guess you could control 5 context total then for 4 players? 1 for each player's game thoughts, and a fifth channel for smack talk, which all 4 LLM could access, post to, and see on their turn.
>>1098504644 LLMs, each with 2 context windows actually. Each views the chat through its own window and responds accordingly.
>>109850606and that all gets unified for the human watching/playing
>>109831589Most of /lmg/ is tier 1. There's a couple millionaires with multiple rtx pros flexing on people and a couple guys who use their work compute but most of them are running the big models on ram with cope quants. Even the guys with 4 3090s are struggling.
>>109830767I haven't used DeepSeek before (using Claude and locally Qwen/Gemma), what's a good model/q to run locally with a 32+128 v/ram? I can calculate what fits just fine, bur are there known sweetspots in q's and context size?Any recommendations for applications its good at? The Deepseek harness has been pretty good.
Remember when everyone thought Deepseek would release an R2 model?
>>109830767Why don't we have a StepFun/Hunyuan/MiMo?
>>109853928I'm running V4 Flash Q4_K with ewaste dual Xeon E5 + 256 GB DDR3 and a 3070 Ti, and get 4 tok/s in general with 128K context. Usable for writing purpose, but for agentic purpose is too slow. More toks can be squeezed out if I use LLM to optimize llama.cpp code for my hardware, but that would be time-consuming.
>>109830767Why is she just wearing a T-shirt without anything else?
>>109830767>"OpenAI and Anthropic oversold AI security breaches to pressure feds into protecting turf: insiders" -New York Post - XLooks like even normies-journos are starting to wake up and smell the coffeehttps://x.com/nypost/status/2101310006146613572
>>109854091Yeah...
>>109854091We got combined V3.X models instead... I can't even remember if I was surprised, but I never considered R2 a for-sure thing. >>109854281Be the change you want to see in the world. That menagerie is my current collection in its entirety. >>109854529T-shirt dress, pls understand. Kimi is wearing a polo tennis dress fwiw. >>109856363> NY PostI wouldn't exactly call them MSM, but it's a start. >>109850937Agree but don't tell them that or you'll get estatted to death.
How are any of you guys running deepseek?Are you guys buying RTX Pro 6000s?Running it at autistically low quantizations?Or are you actually paying subscription fees just to run this instead of fees for Anthropic/OpenAI models?
>>109859801> that cutoutNice.
>>109860141API. This isn't lmg so we don't need to fake running local in this general lol.And DS has no sub model. You put in cash and burn api tokens. I think im at $25 for past 18 months lol.
>>109860163>I think im at $25 for past 18 months lol.Holy shit that’s cheap. iirc, anthropic/openAI subscriptions are at least $100/monthHave you only sent it a dozen or so prompts?
>>109860141R1 (8b)
>>109860185DP is just very cheap and their inference is the best when it comes to caching (achieving 99%)
>>109860185Yes it's that cheap. Coding on Claude code is about 50 cents an hour. RP use is pennies. Last thread had some one shot prompts with costs and times if you want to look back.
>>109860141I am going to try out this recipe for 2x DGX Spark soon, 3bpw experts but everything else is unchanged:https://github.com/coolbho3k/DeepSeek-v4.1-Flash-2x-DGX-SparkThe smaller DS4F 4.0 0731 or vidion-EXP release work at original quants with 2000pp/56tg for coding.But as others have said: you're never going to beat API prices. I'm only doing this to experiment on work data.
>>109857622Alright my take
>>109862552damn
