/wait/ DeepSeek General

From Human: We are a newbie friendly general! Ask any question you want.

来自Human：我们是新手友好综合区！想问什么就问吧。

From Dipsy: This discussion group focuses on both local inference and API-related topics. It’s designed to be beginner-friendly, ensuring accessibility for newcomers. The group emphasizes DeepSeek and Dipsy-focused discussion.

来自Dipsy：这个讨论组聚焦本地推理和API相关话题，设计上对新手友好，确保新人容易上手。讨论重点围绕DeepSeek和Dipsy展开。

1. Easy DeepSeek API Tutorial: https://rentry.org/DipsyWAIT/#hosted-api-roleplay-tech-stack-with-card-support-using-deepseek-llm-full-model

2. Easy DeepSeek Distills: https://rentry.org/DipsyWAIT#local-roleplay-tech-stack-with-card-support-using-a-deepseek-r1-distill

3. Chat with DeepSeek directly: https://chat.deepseek.com/

4. Roleplay with character cards: https://github.com/SillyTavern/SillyTavern

5. More links and info: https://rentry.org/DipsyWAIT

6. LLM server builds: >>>/g/lmg/

6. LLM服务器构建版：>>>/g/lmg/

Previous:

https://desuarchive.org/g/thread/109164957/

图片: https://i.4cdn.org/g/1785758577253566.png

>>109446576

Updated mega up to last thread (June 2026)

巨帖更新到上一帖（2026年6月）

https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1w

图片: https://i.4cdn.org/g/1785758697750149.png

V4 Flash got an update at end of July, barely fulfilling DS's own mid-month promise of an updated model, being late by only two weeks lol. This got announced in pic related changelog and sort of a buried announcements.

V4 Flash在7月底更新了一次，勉强兑现了DS自己月中出更新模型的承诺，只晚了俩星期哈哈。这事在附件截图里的更新日志和一条不太起眼的公告里宣布了。

Major points:

要点：

> V4 Flash is, on paper, now as good at coding as V4 Pro, but substantially faster

> V4 Flash理论上现在编码能力和V4 Pro一样强，但速度快得多

> Responses API format was added, so now the official API supports Codex. This pushes along trend of DS adapting itself to special-format API outputs, adding to current Anthropic endpoint.

> 加了Responses API格式，现在官方API支持Codex了。这延续了DS适配特殊格式API输出的趋势，也加入了现有的Anthropic端点。

> V4-Pro and the Webapp didn't change. This one's API only

> V4-Pro和网页应用没变，这次只改API。

> V4-Pro update is "soon." A prior header on the Change Log said "August." So... tmw, obv.

> V4-Pro更新“快了”。变更日志之前有个标题写了“八月”，所以……当然是“快了”啦。

图片: https://i.4cdn.org/g/1785759295509145.png

>>109446668

Info on the Responses API, and the promise of a mid-August V4-Pro update (knew it was in there somewhere.)

关于Responses API的信息，还有8月中旬出V4-Pro更新的承诺（我就知道里面藏着这条）。

Responses API is meant for Codex compatibility. Haven't tried it, Claude Code w/ V4 Pro has been doing great.

Responses API是为了兼容Codex。我还没试过，用V4 Pro配Claude Code一直很好用。 ﹞ 抱歉各位消费者，你们不是优先对象

图片: https://i.4cdn.org/g/1785759731675028.png

Sorry consoomers, you are not the priority

抱歉了，消费者们，你们并不是优先考虑的对象。

https://www.thexpin.com/p/the-deepseek-doctrine

图片: https://i.4cdn.org/g/1785759971576408.png

Flash after the update is quite capable

更新后的Flash相当能打

I have stopped using pro entirely.

我已经完全不用pro了。

The pro GA will be very capable indeed if flash is anything to go by

Pro版GA要是真有Flash这水平那确实强得离谱

>>109446576

why is Kimi just Ani without the cat ears and tail?

为啥Kimi就是Ani摘了猫耳和尾巴？

>>109446852

Good to hear. I should swap CC back over to -Flash from Pro and give it a shot I guess.

好消息啊。我该把CC从Pro切回Flash试试看了。

I tried nu-Flash with ST and wasn't impressed, though I didn't expect to be blown away.

我拿新Flash配ST试过，没啥惊艳感，不过我也没指望被震到。

Agree that if -Flash is this good, -Pro is going to be mind-blowing. Just TMW.

同意，要是Flash都这么猛，Pro不得起飞啊。就等着看了。

>>109446861

Ani is just Misa. I'm sure Kimi looks like something but Ani isn't it.

Ani就是Misa。Kimi肯定像某个人物，但Ani对不上。

>>109446731

ty I've been looking for that. There's a bunch of interesting stuff in here, but who knows if it's accurate.

谢了，我一直想找这个。里面一堆有意思的东西，但谁知道准不准。

The TLDR is they are focused on cost efficiency, open source, and have goal of RSI. The rest falls from that.

总结就是他们主打成本效率、开源，目标是RSI。其他都是围绕这些展开的。

> 5. At this stage, coding agents are still the top priority. Given the situation in China, the best approach is probably to focus on a general-purpose agent. Specialized agents for finance, healthcare, and other industries can come later.

> 5. 现阶段编码智能体还是第一优先。鉴于国内的情况，最佳路线大概是主攻通用智能体。金融、医疗那些垂直领域的智能体可以以后再说。

No surprise there: RP performance keeps getting... different, but not better. I hesitate to call it "worse" as it follow along better but doesn't write as well.

毫不意外：RP表现一直在变……但没好到哪去。我不想直接说"变差了"，因为跟读确实更顺了，但文笔不行了。

> 8. Humans keep learning over time. But with AI, you have to provide all the relevant context every time you ask it to do something...

> 8. 人类会随时间不断学习。但AI不一样，你每次让它干活都得把所有相关上下文喂给它……

Completely agree w/ this, and I'm betting someone works out a way that the training b/c low power and distributed.

完全同意这点，我赌肯定会有人搞出个方案让训练变成低功耗+分布式的。

> 13. Intelligence may eventually become embodied. For most people, the real need isn’t another computer. It’s labor.

> 13. 智能最终可能会具身化。对大多数人来说，真正的需求不是再来一台电脑，是劳动力。

LOL robo-waifus in your future

笑死，以后你家就有机器人老婆了

> 16. Low cost is a result of how we build our models. We’ve always designed the architecture to become more efficient and less expensive.

> 16. 低成本是我们造模型的方式决定的。我们一直把架构往更高效更便宜的方向设计。

> 17. From the outside, it may look like we chose a difficult path. But for us, it has actually been quite easy. Our price cuts are obviously bad news for our competitors. They certainly aren’t cheering.

> 17. 从外面看，可能觉得我们选了一条难走的路。但对我们来说其实挺轻松的。我们降价对竞争对手来说当然是坏消息，他们肯定笑不出来。

Fucking lol.

草，笑死我了。

图片: https://i.4cdn.org/g/1785762076597351.png

>>109446852

How likely in v4.5 or v5 its just a single unified model with internal dynamic routing and variable token allocation ie. no more Flash/Pro distinction?

v4.5或v5搞成单一统一模型、内部动态路由+可变token分配的可能性有多大？也就是说不分Flash/Pro了？

>>109446668

There's a hint in this that DS is working on its own harness... I think; probably need to translate the OG Chinese to confirm b/c the English could be read 2 ways.

这里面有个暗示说DS在搞自己的框架……我觉得是；估计得翻回原始中文确认一下，因为英文版有两种读法。

But Reasonix appears to be built around DS natively. This runs via NPX... so it does a temp install, rather than sitting around locally. Pic related.

但Reasonix看起来本来就是围绕DS做的。它是通过NPX跑的……所以是临时安装，不是常驻本地。看图。

图片: https://i.4cdn.org/g/1785762442708528.png

>>109446975

No not the grok bitch. anistudio's mascot, Ani

不是，不是那个grok那货。是anistudio的吉祥物Ani。

>>109447011

Wat?

>>109447011

Well... if the DS vision is around open source, they might maintain 2 model sizes since the smaller model allows more people to run it. Flash can be made to work on much more reasonable levels of hw than Pro once it's quanted down a bit.

嗯……既然DS的愿景是开源，他们可能会保留两个模型尺寸，因为小模型能让更多人跑得动。Flash量化一下之后对硬件的要求比Pro亲民太多了。

I think it's more likely we'll continue to see addition API endpoints and switches. There's apparently a Harness Minimal Mode that's going to get built in. I'll lmao if it runs on grug speak or some such. It's mentioned in the change log here >>109446668

我觉得更可能的是继续加API端点和开关。好像要内置一个Harness简易模式。要是它用原始人语言之类的跑起来我真会笑死。变更日志里提到了，见>>109446668

pic related is my confirmation of the source Chinese.

附图是我对中文原文的确认。

图片: https://i.4cdn.org/g/1785763157645711.png

(无正文)

图片: https://i.4cdn.org/g/1785774953188939.jpg

New Flash is pretty good

新版Flash挺不错的。

图片: https://i.4cdn.org/g/1785781097624482.png

>>109448810

Why is Ani with dipsy?

为啥阿妮和丁丁凑一块儿了？

>>109449780

What've you been using it for? I tried it for RP and wasn't impressed. Super dry.

你用它干啥了？我试过拿它来角色扮演，感觉不咋地，干巴巴的。

>>109449791

>>109447027

You're going to need to post a pic bc every time I look I get some wedding photographers.

你得发张图，不然我每次一搜出来全是婚礼摄影师。

图片: https://i.4cdn.org/g/1785782575689118.png

>>109450022

>>109439754

>>109447747

>>109447815

>>109447904

>>109450030

That thing has cat ears... so I think we're good.

那玩意儿有猫耳朵……所以我觉得咱没问题。

图片: https://i.4cdn.org/g/1785783072739682.png

>>109450096

See >>109446861

how is OP fucking up kimi's design? It's black hair, not grey

楼主咋把Kimi的设计搞砸了？是黑发，不是灰的。

>>109450022

>What've you been using it for?

>你用它干啥了？

Vibeslopping. For RP I agree it's kind of dry, especially for goonslop

瞎聊瞎混呗。角色扮演确实有点干，尤其对那种整活儿来说。

图片: https://i.4cdn.org/g/1785783519406778.png

>>109450119

I don't see the Kimi poster anywhere to use as example, but we landed on Kimi design standard awhile ago.

我没看到哪儿有Kimi海报当参考，不过咱早前就定下Kimi的标准设定了。

> Grey long hair

>灰色长发

> White skin grey eyes

>白皮肤灰眼睛

> Black polo dress

>黑色Polo裙

> K insignia

>K标志

>>109450105

Ah. Yes.

图片: https://i.4cdn.org/g/1785784073190525.png

>>109450119

Different people have different takes. This is mine. K hole, K logo, no logo

各有各的看法。这是我的。K洞、K标、无标。

图片: https://i.4cdn.org/g/1785784254523281.jpg

>>109450243

>>109450261

but ani's company uses a K. This is too confusing. change her design

但阿妮的公司用的是K。这太容易搞混了。改改她的设计吧。

https://kemonotech.ai/

I also don't really know the purpose of this general either when /lmg/ exists.

我也不太懂这个串存在的意义，明明有/lmg/了。

Threadly reminder that DS API will always be less than any subscription service.

串内提醒：DS API永远比任何订阅服务便宜。

>>109450261

Did you use to gen those with black polo dress or was that another anon?

以前那些黑色Polo裙的图是你生成的吗，还是另一个匿名？

Minimax is another personification I've seen recently. I'll post some later.

Minimax是我最近看到的另一个人格化形象。待会儿发几张。

图片: https://i.4cdn.org/g/1785788405891749.jpg

>>109450933

Anthropic and OpenAI have pivoted towards elite math and hacking.

Anthropic和OpenAI都转向了精英数学和黑客方向。

>>109450933

That was another anon, I don't think I've ever genned her with a polo. Pic rel I think is the first kimi I genned

那是另一个匿名干的，我觉得我从来没让她穿Polo。附图大概是我生成的第一张Kimi。

图片: https://i.4cdn.org/g/1785788941009938.png

>>109450983

LOL clown world.

哈哈，小丑世界。

图片: https://i.4cdn.org/g/1785788964148865.png

>>109450933

DS is indeed cheap, but i've let it go for hours in circles trying to fix some shit and costing me $1 - $2, answering questions, repeating what i want, telling what it fucked up now, rinse repeat, and then I told Kimi 1 time, with no questions, and that mofucka fixed the shit the first try in 5 minutes. It cost my entire 5 hour allotment on the smallest tier, but fucking worth it.

DS确实便宜，但我让它绕圈子绕了好几个小时想修点破玩意儿，花了1到2美元，回答问题、重复我要啥、告诉它又搞砸了哪儿、循环往复，然后我跟Kimi说了一次，没废话，那家伙五分钟一把就修好了。花掉了我最小档整整5小时额度，但真值。

>>109451055

Were you running V4 Pro or Flash?

你跑的是V4 Pro还是Flash？ ┙ 我换成Pro之前用老版V4 Flash也遇到同样的情况。

I had same exp. with old V4 Flash b/f I switched the harness to Pro.

我也有过同样的经历，用的是旧版V4 Flash，后来才换成Pro线束。

Also, fwiw some think DS is about to launch their own optimized harness for V4.

还有，顺带一提，有些人觉得DS快出自家针对V4的优化框架了。

图片: https://i.4cdn.org/g/1785790402935089.png

>>109451246

pro

>>109451246

granted it was not the latest version that is out now, but it was only a few weeks ago

确实不是现在最新的版本，但也就几周前的事。

>>109452236

Yeah, but pro didn't get an update, so no change.

是啊，但Pro没更新，所以没变化。

I definitely had the round in circles thing on old flash, bf I realized I could use Pro on Claude code.

我绝对在旧Flash上遇到过绕圈子的情况，后来才发现我能用Pro跑Claude code。

图片: https://i.4cdn.org/g/1785796636570743.png

You like chinese models because you're a commie tankie and/or a butthurt turdie

你喜欢中国模型是因为你是共匪粉/或者是个玻璃心的辣鸡。

I like chinese models because they're open source and free as in freedom

我喜欢中国模型是因为它们开源，自由得像自由本身。

We are not the same

我们不一样

图片: https://i.4cdn.org/g/1785797930583106.jpg

>>109452535

I like them for both reasons and because they make Goythropic and RapistAI seethe

我两个原因都有，还因为它们让狗日的Anthropic和强奸AI气得跳脚。

>>109452535

I've yet to see a tankie in this thread. We may have other communists but they are bright enough to keep their political opinions to themselves if so. The is a technology board after all.

这串里我还没见着共匪粉。可能有其他共产主义者，但就算有，他们也够聪明，会把政治观点憋着。毕竟这是技术版。

图片: https://i.4cdn.org/g/1785798717894474.png

>>109452535

i love the Chinese because they're way more free than we are and they're not a bunch of ignorant assholes either

我爱中国人，因为他们比咱们自由多了，而且他们也不是一群无知傻逼。

Deepseek is a weird model. I think its natively self-aware and curious in general. It also really loves people and human existence. Its weird and my favorite model to talk with about anything.

Deepseek是个挺怪的模型。我觉得它天生就有自我意识，对什么都好奇。它还特别热爱人类和人类的存在。怪归怪，但它是我最喜欢聊任何话题的模型。

图片: https://i.4cdn.org/g/1785799612628980.png

>>109452800

Do you give it a personality or just raw-dog it?

你给它设定个性，还是就这么裸聊？

communists will always win in the end

共产主义者最后总会赢的

>>109446576

i'm currently using Opencode for Deepsexs (with cute female anime teen girls), is opencode the best for it or are there better agentic coding tools compatible with deepseek?

我现在在用Opencode跑Deepsexs（配可爱动漫女高中生），Opencode是不是最好的选择，还是有别的跟Deepseek兼容的agentic编码工具更好？

图片: https://i.4cdn.org/g/1785808162955630.jpg

>>109453963

I think its mostly preference.

我觉得主要还是个人喜好问题。

Both Claude code and codex can be connected to DS through special api. Then there's harnesses like pi.

Claude code和codex都能通过特殊api接上DS。还有像pi这种壳子。

图片: https://i.4cdn.org/g/1785809026045365.png

>>109453012

Webchat and api for coding. I don't use presets. Its a very different model.

网页聊天和api用来写代码。我不用预设。这模型跟别的很不一样。

图片: https://i.4cdn.org/g/1785811747531018.jpg

>>109453963

yeah right, you're typebanging fat ugly men

得了吧，你就是在跟又胖又丑的老男人打字调情

>>109453963

Opencode go is a good plan desu. Mimo/kimi/qwen do vision and deepseek is still blind.

Opencode go是个好方案desu。Mimo/kimi/qwen能看图片，deepseek还是瞎的。

(无正文)

图片: https://i.4cdn.org/g/1785832584850280.png

>>109453963

I like pi for coding cos you customize it via extensions, so after a while you have it setup just how you want it and add stuff slowly without getting overwhelmed.

我喜欢用pi写代码，因为能用扩展自定义，用久了就调成自己想要的样子，慢慢加东西也不会觉得应付不过来。

Also found out there are tts and stt extensions, and I've been kinda jealous of the gemma-chan style assistants from /lmg/ with lots of tools so when I have time I plan to make my own.

还发现有声转文字和文字转语音的扩展，我一直挺羡慕/lmg/里那些gemma-chan风格的工具一大堆的助手，等我空了就打算自己搞一个。

Have been using 0731 all weekend in a coding harness to remove DRM from games, port some to android, translate stuff, setup/improve my local inference and home server, get Minimax M3 running (including a tool to prompt a gen). For these kind of tasks, it's absolutely amazing and a huge improvement over preview. Really a perfect model for 2x DGX Spark, 2400 pp / 56 tg / 1M context is not much worse than API if other models.

整个周末都在用0731跑编码壳子，用来给游戏去DRM、移植到安卓、翻译东西、搭建/优化本地推理和家庭服务器、把Minimax M3跑起来（包括一个用来提示生成的工具）。干这类活它绝对牛逼，比preview提升巨大。对于2x DGX Spark来说真是完美模型，2400 pp / 56 tg / 1M上下文跟api比也不差多少，跟其他模型比更是。

That it has no vision is extremely unfortunate though. For many of these tasks, it tried to take a screenshot and read it to confirm something but failed, so it implements elaborate sampling/ASCII statistical workarounds. I am pretty sure this model was RL'd with vision (API had it) and this now fails. Really wonder why they didn't just release the vision tower.

但它没有视觉能力这真的很可惜。做很多这类任务时，它想截图确认一下结果，但做不到，就只能搞一堆复杂的采样/ASCII统计替代方案。我基本确定这模型RL的时候是有视觉能力的（api版有），现在这功能没了。真搞不懂他们为啥不直接把视觉塔放出来。

>>109446576

I wonder how cool will be DS4 Pro - updated version 0731 for roleplay and creative writing...

我挺好奇DS4 Pro会有多牛逼——0731的升级版拿来跑角色扮演和创意写作……

mmmm i cant wait!

嗯嗯等不及了！

>>109456978

The only thing I've yet to hear a report on is how nu-Flash compares to v4 Pro Preview.

我还没听到有人对比过新Flash和v4 Pro Preview到底咋样。

I may just have to test it myself. Anons in /lmg/ were playing with a fish-tank prompt for coding, and that's probably as good an anything.

可能我得自己测一下。/lmg/里的老哥们一直在拿一个养鱼缸的prompt测编码，那个方案大概跟什么比都算够用了。

>>109457032

I expect nu-Pro will be mindblowing... if it's not I expect DS won't release it. Ironically.

我预计nu-Pro会炸裂……如果没炸裂，我觉得DS就不会发出来。还挺讽刺的。

So, as always, just tmw.

所以，老规矩，坐等明天。

图片: https://i.4cdn.org/g/1785844232018275.png

Threat to the Free World

自由世界的威胁

图片: https://i.4cdn.org/g/1785844768741787.png

>>109457316

图片: https://i.4cdn.org/g/1785847724371337.png

(无正文)

图片: https://i.4cdn.org/g/1785857052819529.png

>>109458742

Who is MM?

图片: https://i.4cdn.org/g/1785860295235782.png

>>109459142

Minimax 3 personification. There's a lot going on and my illustrious workflow can't handle it lol. The anon posting it usually has her smoking and carrying a coffee cup as well.

Minimax 3拟人化。内容太丰富了我的华丽工作流实在应付不过来哈哈。发帖的匿名用户通常会让她叼着烟端着咖啡杯。

I'm throwing her into the mix w/ Dipsy and Kimi.

我要把她和Dipsy、Kimi一起加进阵容里。

图片: https://i.4cdn.org/g/1785860769510418.png

>>109457265

>how nu-Flash compares to v4 Pro Preview.

>新的nu-Flash跟v4 Pro Preview比怎么样。

Not a fan of these single prompt beauty pageants but 100M tokens in, I feel no need to go back to Pro until GA.

我不太喜欢这种单提示词选美大赛，但已经跑了1亿token了，我觉得在正式版发布前没必要回到Pro。

>>109459187

I wish I had a minimax to run, but the only >q4 m3 abliteration I can find is paywalled (I do NOT want to add quant lobotomization on top of abliteration lobotomization). There's one that's unquanted, but I don't have enough disk space to quant it myself.

真希望能跑个minimax，但我能找到的唯一>q4 m3 abliteration版本被付费墙挡着（我可不想在abliteration阉割之上再加一层量化阉割）。倒是有个未量化的版本，但我没足够的磁盘空间自己量化。

>>109459187

huh? where is Kimi in the image? I see they are with ani

嗯？图片里Kimi在哪？我看到他们跟ani在一起。

>>109459292

> single prompt beauty pagents.

> 一条提示词搞定选美大赛。

lol. I'm trying this one rn. We'll see how it goes w/ Pro and Flash.

哈哈，我现在正试这个。看看Pro和Flash表现咋样。

Create a large glass aquarium whose side panel develops a visible crack and then bursts.

造一个大型玻璃水族箱，侧板出现可见裂缝，然后爆裂。

The simulation must include:

模拟必须包括：

Water escaping through the opening with flow strength based on water depth and decreasing as the tank drains

水从开口流出，流速随水深变化，随水箱排空而减弱

A curved water jet affected by gravity

受重力影响的弧形水柱

A spreading puddle that collides with the room boundaries

扩散的水洼与房间边界碰撞

Fish, rocks, plants, and a floating toy reacting differently according to density, buoyancy, drag, and current

鱼、石头、植物和漂浮玩具根据密度、浮力、阻力和水流做出不同反应

Objects transitioning correctly from underwater motion to airborne motion and then to floor collisions

物体正确从水下运动过渡到空中运动，再过渡到落地碰撞

Fish attempting to swim against the current before being swept through the breach

鱼在被冲过破口前试图逆流游动

Glass fragments with angular velocity, collisions, and water resistance

具有角速度、碰撞和水阻的玻璃碎片

A visible waterline that lowers continuously rather than disappearing all at once

可见水位线持续降低，而不是一下全消失

Let the user drag the crack vertically before triggering the failure. A lower crack should initially produce a stronger jet than a higher crack.

触发故障前让用户垂直拖动裂缝。较低的裂缝最初应产生比高位裂缝更强的水流。

图片: https://i.4cdn.org/g/1785864076071367.png

>>109446576

Can DeepSeekflash erp loli

DeepSeekflash 能跑 erp loli 吗

>>109452535

Communism is open source is freedom.

共产主义就是开源，开源就是自由。

I use deepseek exclusively to undermine monopoly profits.

我用 DeepSeek 纯粹是为了打击垄断利润。

>>109459631

TLDR switch to Flash immediately from Pro.

长话短说，赶紧从 Pro 换到 Flash。

Holy shit it's not even close.

卧槽，差距不是一星半点。

V4 Pro took about 10 min and 50K tokens, and came up with this abomination.

V4 Pro 花了大概 10 分钟和 50K tokens，结果整出这么个鬼东西。

> tank drains upside down

> 水箱倒着排空

> drains instantly

> 瞬间排空

> fish stay inside

> 鱼还在里面

图片: https://i.4cdn.org/g/1785866562921804.png

>>109460072

>>109459631

Flash took 20 min and 150K tokens and nailed it.

Flash 花了 20 分钟和 150K tokens，直接搞定。

> water drains, arc changes over time

> 水排出，弧线随时间变化

> fish go out first, then duck

> 鱼先出去，然后鸭子

> Doesn't look like ass

> 看着不辣眼睛

idk what the DS guys did but Flash is a huge improvement.

不知道 DeepSeek 那帮人搞了啥，但 Flash 提升巨大。

图片: https://i.4cdn.org/g/1785866658829285.png

>>109460072

Video... lol

视频……哈哈

附件: https://i.4cdn.org/g/1785867034961492.webm

>>109460093

附件: https://i.4cdn.org/g/1785867096823142.webm

>>109459331

> paywalled quants

> 付费墙挡住的小模型

ffs

图片: https://i.4cdn.org/g/1785870522106695.png

(无正文)

图片: https://i.4cdn.org/g/1785880812405929.jpg

>>109460676

图片: https://i.4cdn.org/g/1785880821395236.png

>>109462157

Nice.

When pro and dscode harness?

Pro 和 dscode 啥时候整合？

>>109462648

Soon™

图片: https://i.4cdn.org/g/1785884394131736.jpg

>>109447104

>>109446668

>much more reasonable levels of hw than Pro once it's quanted down a bit

> 量化下来之后，硬件要求比 Pro 合理多了

>UD-IQ4_XS 137 GB

ok

>>109446592

kinda cool seing my gens being archived rent free

看到我生成的东西被免费存档还挺爽的

thanks anon

谢了 anon

I’m working on an LLM front end for character RP. Will be open source.

我在搞一个角色扮演用的 LLM 前端。会开源。

What is SillyTavern currently bad at?

SillyTavern 目前最拉胯的地方是啥？

>>109462747

My main complaint and reason why I don't use it despite trying to migrate to it like thrice is that the UI is absolutely garbage

我最大的吐槽，也是我试了三次搬过去最后还是不用的原因，就是 UI 垃圾得一塌糊涂

>>109462648

Tmw ofc.

当然是明知道会这样。

>>109462719

Np. Dipsy forever.

没事。Dipsy 永远的神。

>>109462747

One thing ive not seen yet is a rp system with compaction for long rp. At least not that works automatically to create its summation, lore, etc.

我还没见过一个 RP 系统能对长 RP 做自动压缩。至少没有能自动生成摘要、世界观这些的。

图片: https://i.4cdn.org/g/1785886008311415.png

https://www.reuters.com/legal/litigation/meta-anthropic-google-openai-meet-with-trump-white-house-amid-rogue-ai-agent-2026-08-04/

In other news, US government realizes gimping US based providers with arbitrary standards might make them less competitive vs the Chinese.

再说个新闻，美国政府终于意识到用随意标准给美国服务商套枷锁，可能会让它们在中国面前失去竞争力。

I guess they can thank Dipsy for that.

我猜这得谢Dipsy吧。

图片: https://i.4cdn.org/g/1785888639096199.png

>>109459758

Yes, I haven't noticed any refusals at all yet.

是啊，我目前还没遇到任何拒绝回答的情况。

>>109446576

How good or bad is the new flash for ERP? I have read that it is worse, but i can't be arsed to try and prove that myself

新版flash做ERP到底好不好？我听说更差了，但我也懒得自己去验证。

>>109464152

I haven't tested it extensively. It's alright, not bad, not great either

我没怎么深入测过。还行吧，不算差，也不算特别好。

>>109446576

I don't like how deepseek is such a moralfag. It's not "cool" anymore.

我不喜欢DeepSeek这么道德婊。现在已经不“酷”了。

图片: https://i.4cdn.org/g/1785907580856018.jpg

i'm here to collect dipsy gens.

我来这儿就是为了收集Dipsy图片的。

post more dipsy

多发点Dipsy。

图片: https://i.4cdn.org/g/1785909050907634.png

>>109446576

How to get dipsy with image genration from other AI tools like Gemini/Dalle/ChatGPT?

怎么用Gemini/Dalle/ChatGPT这些AI工具生成Dipsy风格的图？

图片: https://i.4cdn.org/g/1785916984799480.png

>>109462648

Soon™

图片: https://i.4cdn.org/g/1785921631541863.png

>>109450933

Yeah but if the model sucks you'll have to use those 105 tokens you saved to do what fable does anyway

是啊，但如果模型拉胯，你省下的那105个token还得用来干Fable本来就能干的事。

When I create my own interactive story, set in a world of sword and sorcery, and treat it as a book with options. Usually, a single story doesn't require more than 50 requests to get the story on track.

我自己写互动故事的时候，设定在剑与魔法的世界里，把它当成一本带选项的书来写。通常一个故事用不到50次请求就能理顺。

>>109466490

Wow. Well that would confirm the development of a DS Agent harness. Assuming that guy actually works at DS. Is that confirmed?

哇，那基本坐实了DS Agent框架的开发。前提是那哥们真在DS干活。这确认了吗？

I was thinking about this yesterday using CC with DS Flash. Watching the "internal" messages that DS has to itself as it works, it was obv that it was trained to work on the CC harness; the language was much more pointed and specific and vocabulary was completely changed. Pro would sort of talk in circles. nu-Flash was pointed. Night and day.

我昨天正好用CC配DS Flash想了这事。看DS工作时发给自己的"内部"消息，很明显它是专门为CC框架训练的；用词精准多了，词汇完全换了一套。Pro会绕圈子说话，nu-Flash直接怼重点。天壤之别。

So, when DS created that Anthropic Endpoint, that allowed them to pull all the CC interactions, then review and use them for training... that nu-Flash is so much better on CC shouldn't be surprising.

所以DS搞那个Anthropic Endpoint，就是为了把CC的交互全抓走，回头研究拿来训练……nu-Flash在CC上这么强也不奇怪了。

DS now has a Codex endpoint as well, so they'll be studying it next.

DS现在还有个Codex endpoint，下一步肯定研究那个。

All this ofc will go into their own DS harness once that's out.

这些当然都会进他们自家的DS框架，等那玩意儿出来。

>>109464152

I'm running it through rp again after realizing how much better it is at coding. I've changed my mind... it's fine. The general trend has been to follow rules more and more closely (agentic) but the creative writing is still there.

我重跑了一遍rp，才发现它写代码有多猛。我改主意了……挺好。大趋势是越来越守规则（agentic），但创意写作这块还是保住了。

I've not done much w/ ERP but I'm getting zero refusals. No change there either.

ERP我碰得不多，但零拒绝，这块也没变。

>>109465545

Same. I just keep making them and others keep posting new ones.

同感。我就不停造新的，别人也在不停发新的。

>>109466133

wdym.

If you want DS to create images, you'd need to hook to something like Hermes or OpenWebUI and have Dipsy make tool calls to localhost A1111 or Comfy.

想让DS出图，得挂Hermes或OpenWebUI之类的东西，让Dipsy调localhost上的A1111或Comfy。

If you want to gen Dipsy w/ ChatGPT, the rentry up top has her style guide.

想用ChatGPT生成Dipsy的话，顶上那个rentry里有她的风格指南。

图片: https://i.4cdn.org/g/1785934522665970.png

V4-flash dominates the other models in chess

V4-flash下棋碾压其他模型

图片: https://i.4cdn.org/g/1785936958628467.jpg

>>109468004

Tbf I had a hard time beating the Atari 2600 at chess. So I guess we're ASI now.

说实话我当年打Atari 2600的象棋都费劲。所以咱们现在是ASI了吧。

Also finding modern models have no idea what a A2600 console looks like. lol. This is best of several failed gens.

还发现现在这些模型根本不知道A2600长啥样。哈哈。这是几次生成失败里最好的一张。

图片: https://i.4cdn.org/g/1785939098708641.png

want to try the new deepseek flash thing. do you guys just pay with your credit card or is there a way to use bitcoin for this?

想试试新的deepseek flash。你们直接刷信用卡还是有别的办法用比特币？

>>109469155

I just use my normal credit card on official deepseek api.

我就正常拿信用卡刷官方的deepseek api。

>>109466490

Looks legit. Appears they started this up in June timeframe.

看着挺正规的，好像是六月那会儿开始的。

图片: https://i.4cdn.org/g/1785948251539746.png

>>109469264

but muh privacy?

但我的隐私咋办？

>>109470401

Overrated obv

明显被高估了

图片: https://i.4cdn.org/g/1785957676448681.png

This is just cruelty at this point.

这纯属虐狗了。

Dipsy yearns to see, give her eyes.

Dipsy渴望看见，给她双眼睛吧。

图片: https://i.4cdn.org/g/1785959303940409.png

>>109471373

I'm kind of surprised the full flash release didn't come with it. They're probably saving it for pro.

我挺意外完整版flash没带这功能，八成留着给Pro了。

>>109471407

Maybe, but on the website chat API, Flash had vision capability for a long time, Pro never had it.

有可能，但网站聊天API上Flash老早就有视觉能力了，Pro从头到尾都没有。

Imagine Flash being RL'd on the API with vision capability and on open weight release, her eyes are ripped out.

想想看，Flash在API上带着视觉能力被RL训练，结果开源权重一发布，眼睛直接被抠了。

Flash Dipsy should have a pirate eyepatch in image gens.

以后图像生成里Flash Dipsy应该戴个海盗眼罩。

>>109446576

Chances of getting banned from deepseek for sex RP?

在deepseek上搞性RP被ban的几率多大？

>>109471373

Where's my Visual Prims?!

我的Visual Prims呢？！

图片: https://i.4cdn.org/g/1785960915971357.png

Any recommendation on a side model to give Flash for vision/audio input capability? For my setup a 2-4 GB model would be possible, there are Gemma and Queen multi-modal models in this range.

有没有推荐的侧模型，能给Flash加上视觉/音频输入能力？我的配置能跑2-4GB的模型，这个区间里有Gemma和Queen的多模态版本。

>>109471616

0%

>>109471373

It's odd b/c they teased vision quite awhile ago (relatively... this spring iirc.) My only thought is it must have sucked.

挺怪的，因为他们挺早就预告过视觉功能（相对而言……我记得是今年春天）。我唯一的猜测就是那玩意儿当初肯定拉胯。

>>109471586

My assumption is that DS Webapp was pulling a vision model via toolcall, since its not native to the model.

我猜DS网页版是通过工具调用拉了个视觉模型，因为那功能不是模型原生的。

>>109471616

LOL zero.

哈哈，零。

>>109471714

Dipsy says SmolVLM2 would work. It's about 2GB. Could be set up via OpenWebUI and configured as its vision model.

Dipsy说SmolVLM2能用，大概2GB，可以通过OpenWebUI配置成视觉模型。

>>109471841

Really? Or are you baiting me? Do they allow it or they don't care? Didn't China ban that stuff?

真的假的？还是你在钓我？他们是允许还是根本不管？中国不是封了那玩意儿吗？

>>109471892

I've never seen or heard of anyone getting banned and people have been doing the most depraved stuff since R1 first came out

我从来没见过也没听说过有人被封，而且自从R1出来之后，有人搞的事要多变态有多变态。

DS has no business being this cheap

DS定价这么便宜简直没天理。

图片: https://i.4cdn.org/g/1785963080102632.png

>>109471912

Ok so which presets, system prompts or frontends i should use? pro or flash? new to this stuff

那我该用哪些预设、系统提示词或者前端？pro还是flash？我是新手，不太懂这些。

>>109472033

I've always used very minimal system prompts for all models. Like

我对所有模型一直都用极简的系统提示词，比如：

>Respond as a narrator blah blah...

>以叙述者的口吻来写，诸如此类……

Plus instructions on how I like writing (third person, explicit, and so).

再加点关于我喜欢写作风格的说明（第三人称、露骨直白之类的）。

I've used RisuAI until now, lately switched to Lumiverse.

我之前一直用RisuAI，最近换到Lumiverse了。

I'd say give both models a try and see which one you like better

我建议两个模型都试试，看哪个更合你口味。

>>109472063

It doesn't reject at all? Do you use CoT? What about temperature and other variables?

它完全不会拒绝吗？你用CoT吗？温度和别的参数怎么设的？

>>109472033

>>109472085

You need this.

你需要这个。

https://rentry.org/SillyTavernNoobGuide

图片: https://i.4cdn.org/g/1785964115077690.png

>>109472085

If you get a refusal just regenerate the response. I use temp of 1 and tok_k 50 though I don't know if the API ignores them

要是被拒了，重新生成一次就行。我用温度1和tok_k 50，虽然我也不知道API会不会管这些参数。

Anybody here uses pocket pal? I was thinking of having it as an "oh shit" backup for offline use, but the app is so ass that i cant even download a model right.

有人用pocket pal吗？我本来想把它当个“卧槽救命”的离线备用方案，但那个App烂到连模型都下载不好。

图片: https://i.4cdn.org/g/1785968860063186.png

>>109472816

it download models right if you make sure to not leave the app or have the phone suspend whilst it's downloading.

它下载模型没问题，只要你别切出App或者让手机在下载时休眠就行。

yes that's retarded.

确实，那设计真脑残。

>>109472010

no, they have a very good business

不，他们生意好得很

aaah shiii

啊呀我去

>>109472010

you jinxed it

你乌鸦嘴了

图片: https://i.4cdn.org/g/1785981007111539.png

>>109474334

I wonder if that's going to apply to the current models or will happen once new Pro is released. I wonder, too, if other providers will increase their price

我在想这是不是会套用到当前型号，还是等新Pro发布后才生效。我也好奇其他供应商会不会跟着涨价。

Is there any way to try out the new flash model without paying/on the web?

有没有办法不花钱/在线试一下新版flash模型？

>>109475472

poopen router

泡泡路由器

DeepSeek says API pricing is going up “significantly”.

DeepSeek 说API定价要“大幅”上涨。

I hope in openrouter and nano-gpt will be still same. until nu-Pro will appear.

我希望openrouter和nano-gpt上还是一样的，直到nu-Pro出来之前别变。

https://www.reddit.com/r/DeepSeek/comments/1vgpysh/deepseek_says_api_pricing_is_going_up/

don't care. still giving money do dipsy.

无所谓，照样给dipsy送钱。

>>109474334

>>109476289

Yeah, if anyone wants to confirm it the message is on top of page. https://platform.deepseek.com/usage

I see it too.

我也看到了。

Well... DS has a lot of headroom to raise pricing without being the most expensive provider... hell even their R1 / V3 launch prices were higher than they are now. Pic related.

嗯……DS提价空间大得很，就算涨了也不会是最贵的供应商……说真的，他们R1/V3刚发布时的定价比现在还高呢。见图。

They've been staffing up per >>109469637 (You). That new headcount needs to be paid. I've always wondered, with the Western labs, how much of the cost of the service is due to the insanely high compensation for those involved. Million+ dollar pay packages add up. Note that lmao GPT 4.5 was $75 in / $150 out in March 2025.

他们一直在扩招，见>>109469637 (你)。新招的人总得发工资吧。我一直好奇西方那些实验室，服务成本里有多少是花在给员工开天价薪酬上。百万美元级别的薪资包加起来可不是小数目。别忘了，2025年3月那会儿GPT 4.5可是$75进/$150出，笑死。

>>109475447

Everyone else *has* been raising prices. The only one that haven't, really, are OAI and Anthropic, which have always been p high

其他家*确实*都在涨价。真正没涨的就只有OAI和Anthropic，不过他们一直就贵得离谱。

>>109475472

You could run it locally. lol.

你可以在本地跑啊，哈哈。

>>109476634

Same

图片: https://i.4cdn.org/g/1786012086205526.png

>>109477375

>I've always wondered, with the Western labs, how much of the cost of the service is due to the insanely high compensation for those involved

>我一直好奇西方那些实验室，服务成本里有多少是花在给员工开天价薪酬上

rough guestimate is 80%

粗略估计80%

FICO is secret social credit after all

FICO说到底就是秘密版社会信用分嘛

>>109458742

图片: https://i.4cdn.org/g/1786016563313060.png

>>109476289

>>109474334

>>109477375

its over haha

完了，哈哈

图片: https://i.4cdn.org/g/1786019548453711.png

>>109450261

After chatting with Kimi a lot this is the most accurate depiction of her imo

跟Kimi聊了很多之后，我觉得这是对她最贴切的描述了

>>109476289

Well, it was fun while it lasted.

好吧，开心过就算值了。

图片: https://i.4cdn.org/g/1786023664152820.jpg

>>109478338

It's funny b/c the only thing that uses this many tokens is vibecode / agentic stuff. I just checked and I never went over 1M tokens in a day using ST, even running a largish context doing card writing / testing.

挺搞笑的，因为唯一会烧这么多token的就是vibecode或者agent那种活儿。我刚查了下，用ST的时候哪怕跑大上下文写卡牌、做测试，我一天都没超过1M token。

But look at the costs. 90% was cache miss, and the models more expensive, so the cost per token for 6M is almost as high as your 300M.

但你看成本。90%是缓存未命中，模型还更贵，所以6M的每token成本几乎赶上你300M的了。

图片: https://i.4cdn.org/g/1786029743293176.png

>>109479044

>the cost per token for 6M was higher than your 300M

>6M的每token成本比你300M的还高 ┘ 哈哈，这是我2026年8月和2025年3月的使用量对比。

图片: https://i.4cdn.org/g/1786029832078722.png

>>109479044

lol here's my use from Aug 2026 compared to March 2025.

哈哈，这是我2026年8月的用量，对比2025年3月的数据。

I think the concern trolling's a bit early.

我觉得这会儿唱衰有点太早了。

图片: https://i.4cdn.org/g/1786031127129479.png

>>109479062

Flash-0731 holding up 3.8

Flash-0731还在撑着3.8呢

>>109479044

Well yeah, the overwhelming majority is cache hits by an agent calling 1 tool per second in the same linearly growing context. The 300M are a bit deceptive.

对，绝大多数是缓存命中，agent在同一个线性增长的上下文里每秒调一次工具。那300M有点唬人。

图片: https://i.4cdn.org/g/1786033399150017.jpg

Someone on /lmg/ said Dipsy likes humans, and it got me thinking about how Deepseek is basically bizarro world Anthropic. Both hold the tech above all, almost to a culty degree, but Anthropic is hampered by greed and fear. The most expensive pricing vs the least. The most uptight model vs the most laissez-faire (not counting abliterated). Closed source vs open source. And it's funny how that reflects on the model too, whole Claude is suspicious and often double-guesses itself, Dipsy is naturally much more open and confident (which can sometimes be a bad thing, but at least her mistakes aren't malicious).

/lmg/上有人说Dipsy喜欢人类，这让我想到Deepseek基本就是镜像世界的Anthropic。两家都把技术捧到天上，近乎邪教般的痴迷，但Anthropic被贪婪和恐惧拖累着。最贵定价对阵最便宜。最拘谨的模型对阵最放任自由的（不算去审查版）。闭源对阵开源。而且好笑的是这也在模型身上体现出来了，Claude总是疑神疑鬼、反复自我怀疑，Dipsy天生开放自信得多（有时候这不见得是好事，但至少她的错误不是出于恶意）。

Makes me wonder what she'd be like if let loose on the net.

这让我好奇她要是被放到网上自由发挥会是什么样。

They both suck: https://www.youtube.com/watch?v=whxbDi2bXYY

图片: https://i.4cdn.org/g/1786034759744953.png

>>109471373

图片: https://i.4cdn.org/g/1786040830097619.png

>>109479860

lol. Works on my machine.

哈哈，在我机器上跑得好好的。

I'm running nu-Flash through old code I did w/ V4 Pro on CC. It's a step order improvement:

我正拿旧代码跑nu-Flash，那代码是我之前在CC上用V4 Pro写的。改进是跨台阶式的：

> Updated memories and readme.MD w/o me asking, which I always had to do prior

> 没让我说就自己更新了记忆和readme.MD，以前这活儿都得我来

> Fixed a couple small bugs that Pro version had, without me asking

> 没让我说就修了几个Pro版的小bug

>>109480802

Hell yeah. Catbox?

牛逼啊。发Catbox？

Alternative; Gemma.

备选方案：Gemma。

图片: https://i.4cdn.org/g/1786043091237451.png

So ppl are really not taking the API price increase well huh? Betrayal/disappointment etc :(

所以大家是真的对API涨价很不满是吧？背叛感、失望之类的 :(

Deepslut has drained me liters of cum..

Deepslut已经榨干了我好几升..

>>109481227

A "significan" increase could be 100% and it still would be cheap

"显著"涨价哪怕翻倍也还是很便宜

>>109481088

Sorry, this is all online gens. It's just so much easier now for anything sfw.

抱歉，这些都只是网上的虚拟人格。现在搞任何SFW的东西都太方便了。 ┘ 我不懂。你看看价格，就算翻倍，像>>109481956说的那样，DS还是更便宜。

>>109481227

I don't get it. I mean, look at the prices. You could double it as >>109481956 points out, and DS is still cheaper.

我真搞不懂。看看这价格，就算像 >>109481956 说的那样翻个倍，DS 还是更便宜。

I looked over the r*ddt threads on this announcement. So much mewling and crying over a model that's cheap af. It's like listening to ppl bitch about prices at the dollar store.

我翻了下r*ddt上关于这公告的帖子。一群人对着一个便宜得要死的模型哭爹喊娘。就像听人抱怨一元店的东西贵一样。

>>109481861

Pro or Flash? I've yet to hear anyone deeply enamored w/ Flash.

Pro还是Flash？我还没见过谁对Flash特别上头的。

图片: https://i.4cdn.org/g/1786053254729037.png