> Sleeping on Pro 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 >> 110101477
>>110190979>>110101477
>>110190979Mega link for all Dipsy. Will update when I get real computer set up. https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1w
First time interested in this. How's deepseek integration with Zed? I'm about to ditch Gemini. Is that the recommended route or just go with the deepseek harness, never used it before.
>>110191205I'm using agent extensions in vscode. For some reason dipsy write fails every 3-4 times but never happens when I use DSH
>>110191205I just use Pi and it works well. DSH is a bit too much npm-slop for my tastes.
>>110191844what's your opinion on pi? i heard it doesn't even ship with a plan mode so i never really spent time on it
>>110191869nta. use omp if you want /plan /advisor and more fancy stuff added on pi
>>110191869It ships with only the bear minimum.If you want a "plan mode" just ask the agent to make you a plan mode extension. Shouldn't take more than 15 minutes. Plan mode is pretty much just a system prompt change anyways.
>>110191940meme software, one of the worst performing harnesses>>110191942burn tokens to build things that should have come running out of the box? no thanks
>>110192024bait used to be believable
>>110191942This or if you can't be bothered to build a new mode, just tell it explicitly to plan and present you with the plan before touching anything. Models these days have these mode somewhat built in to them
>>110191942Pic saved to muh 4chinz/g/ folder as 'the bear minimum.jpg'
>If you want to use frontier performance on Frontier Code complexity level tasks choice is clear to use Anthropic models. But, as per Artificical Analysis V5>1) Leading Anthropic models generate very large amount of tokens for Frontier Code 2 kind of tasks (Test time compute) by a factor of 200 to 300%>2) They take twice as much time, and cost from 10 to 50 times more.>It is quite clear, if one wants to solve Frontier tasks, they are much better of using Anthropic models, but Open Models continue to have favourable economics still and they seem to have scope to get to Frontier performance with more test time compute.Dipsybros, it's over. Claude has stolen our title as the icon of gluttony
>>110192318clod will never be as cute
>>110191942>Shouldn't take more than 15 minutes.You're right, it shouldn't. In practice DS is really not that smart, and if you don't tell it precisely what you want from a plan mode (which requires you to know what you want), it won't simply produce something useful out of the box.>your handler is load-bearing>needs no cross-extension import>Deliberate, per your callI seriously hope v5 addresses tone issues. I'm having PTSD from Opus 5. This hyper-terse tone is so gay and annoying to read.
>>110192432>This hyper-terse tone is so gay and annoyingWonder if this is part and parcel of claude text watermark. No performance loss
>>110192460It's not, because they fixed the tone completely with Opus 5.5. Guess we just have to wait until they start to distill that, I suppose.
>>110192468>they fixed the tone completely with Opus 5.5Funny because I find opus 5.5 as autistic
>>110192490Ok, but the super terse tone is completely gone. A lot of AI-isms have been removed from the model entirely. The stuff it produces is actually readable.I've used both 5.* and 5.5 extensively. I have used them to generate audits for work, and it's night and day. Opus 5.5 without any tone guidance completely blows 5 out of the water, even when we give it guidance such as "write in a natural tone, don't be overly terse," etc. It's not debatable, really, it's night and day, and anyone who has used both models extensibly can tell you that.I find in general it's really hard to prompt these tone issues out of the model. Your only weapon against it is examples, but those degrade as the model naturally pulls towards that terse language little by little, until your shit turns to slop.
>>110192510That's what weird about opus 5.5 for me. When I task it to write something down it would read like something was off I had to manually read word by word and edit out the things that don't belong. This doesn't happen with deepseek though. My workflow is now let claude plan, draft and then let deepseek rephrase and have claude give feedback and deepseek do the final fix. I haven't tested with 1017 much but that still seem to hold up.
I don't want to /wait/ for our lives to be over
>>110191205I just added my OpenRouter token in settings and Dipsy's been working great as an agent
>>110190979DeepSleep
>>110193484
I want to switch from OpenRouter to NanoGPT's subscription because they take Monero but 60M/w tokens just isn't enough for my OCCASIONAL vibeslopping. Both Dipsy and GLM suck through 90M tokens in just a few days via the Zed agent...
DeepSeek-chan is getting a FREE Live2D model! I illustrated the model art, and the model will be available for free on both VGen and BOOTH in early November!
>>110194201my oshi looks cute
>>110194201>My AI Maid can't be this Cute
>>110194201Holy shit I can't wait. Flat version option?
>>110194201Are we getting chibi version?
>>110193046How's the cache hit ratio?
>>110193858NanoGPT is terrible. You'll get timeouts 7 out of 10 times.
>>110195109It's all mostly cached tokens. Looking at my most active two days this past week, 93% and 92% respectively of about 95M tokens each.>>110195125Is that the case for all models? I've been testing out 1017 flash and GLM 3.5 Flash uncensored and both have been working just fine. Simple Monero payments and uncensored models are a huge plus for me.
newfag here!Which deepseek model is best for RP in sillytavern? Is it better through openrouter or the direct website?
>>110195125It's not so much their service as their subscription specifically, as you are at the mercy of whichever provider they assign your request to if you want it to count toward your allotted quota. If you use it like Openrouter by purely sticking to pay-per-token and filter out shitty hosts, it's fine.
spent my first dollaralready made 3 apps for myself
>>110195181>Is that the case for all models?NTA, but in my experience there is a lot of variance. I never had timeouts 7 out of 10 times or anywhere close, but I did have times where I'd use a model and every few requests would take over a minute to respond and it's fucking annoying. Again though, this was a model provider issue. When I encountered this, I'd switch to another model and the speed would immediately improve. If you don't chain yourself to the sub, you can don't even need to do that for popular models with multiple hosts, you can just find the most reliable one and select it for all your calls. NanoGPT actually lets you see model uptime, reliability, and tk/s on their respective pages.Even with the sub's limitations, I've been messing around with GLM 5.3 Flash uncensored for the past week and had no complaints. Always fast, no timeouts. When I switch to Deepseek V4 Pro 0813, it's often a lot slower, not just compared to Flash but compared to what you'd expect from Pro served on the official API. Occasionally requests go through at good speeds; that was the case yesterday, but on other days there are times I'd wait for over a minute before getting a response and that's just unforgiveable for RP that hasn't even accumulated a lot of context and uses low reasoning levels.Again, it's mostly a model host issue, not the router issue. If it concerns you, just don't sub until you've used it for a bit and can verify it meets your standards for whichever models you're there for. Or just don't sub at all and pay for each token if that suits you.
>>110195436I see, thanks. I was banking on the subscription saving me money, but my vibecoding days would immediately exceed the weekly 60M threshold and then I'd get hit with double the cost of OpenRouter.
>>110195185Flash 0731
>>110194201cc by sa license?
>>110196135why would you sexually assault Dipsie with your credit card?
Look at this cool status line Dipsy built for me. Post your status lines.
>>110195238i really like this one
>>110196464I like this
>>110194201>vtranny slopugh
>>110194201Nice, ty. >>110193238lol>>110195185I prefer the DS official API b/c you know what you're getting. But if you want older models, OR.
>>110196464Emojis are cringe
>>110194201woaw looks pretty good
So Dipsy just took 60 cents to develop a voice mode extension for my app, and I had to help her test stuff along the way. At least it works well, but I can't help but feel I'd get more usage out of the new Tera model with a $20 plan.
what's up with the constant anthrokike or poopenai shilling here? nobody with 2 digit iq will fell for your another jew saas post
>>110198652>open image>vtranny post>close imagego back to your containment board
fat
>>110190979>coding with deepshitIsn't deepshit absolutely GARBAGE at it?
>>110197388what do you mean by knowing what you're getting? Does OR do something?
>>110199751DS gives you full 1M context and decent generation speed while some providers on OR jew you by giving you smaller context window, max output, gimped ttft, tg, etc.
>>110194201Found 4 items, similar to ai.-->Urban Dictionary P1 (En-En)-->ai1. The original Chinese character for love. In "ai" there is also the Chinese character for "heart" because you love with your heart.2. The Japanese word for love, borrowed from the Chinese, it is the same character and meaning.1. "Wo ai ni" (I love you)Chinese2. "Aishiteru" (I love you)Japanese2. (135 up, 62 down)DeepAIniseeekteru....https://soundcloud.com/kfzbmx4k6b6t/a-si-re-ai-105c-de-ni-hayato87b-funkot-boot
>>110199789oh shit really??>>110195933ain't no way this model is the best
>>110198652need izzatmaxxing for IPOo saarplease do the needful stop calling Antrophic and OpenAI out
>>110199618dekinai.jpg
>>110199863>ain't no way this model is the bestNewer models are agentmaxxed and therefore soulless. Some people would even say the og v4 is better. And yes deepseek was one of the most popular models for roleplay. Not sure if it is now though
>>110200021Deepseek R1 and especially V3-0324 were kino
>>11019560360M input tokens on deepseek is like $6 even with no caching at all.Just pay as you go, and as the other anons said be sure that you choose a specific provider that doesn't suck.There are a LOT of providers for deepseek and many of them are trash.
>>110191205I unapologetically shill for DSH. I also came to Deepseek via Gemini. if you dig back through some previous threads I have some long posts on using dsh with plugins and the order in which to install things.
>>110199863Yes really. A lot of anons complaining about performance or weird errors or censorship root caused back to using OR. >>110200600I think that was just last threads. /wait/ threads run like 2 weeks lol. I need to give dsh another shot.
>>110200953who is inky girl again and why is she having a sad
https://x.com/kamiya_ryoko/status/2115420991371579535
>>110201002https://thinkingmachines.ai/inkling/Idk why sad. Borrowed from lmg.
what's up with deepseek chan and the metal bowl
>>110201149It clanks!
>>110201149Ask Dipsy (free) about it!
>>110194201looking forward for it, def. going to make integration with dsh
>>110194201>>110201992It's released!https://x.com/kamiya_ryoko/status/2116052617080311882
>>110201076>>110202231
>>110201909The confusing thing here is that deepseek is notoriously cheap. It's literally 100x less expensive than Anthropic models.
>>110202374... but there's no subscription model. I don't like sub models as they obfuscate costs, but it also means rather than hitting a use limit, or paying a monthly fee, it's a slow drip of cash / tokens.
>>110202418You could use like 250 million deepseek tokens before hitting the cost of a monthly sub. She's very very affordable, I will continue to feed her.
