File: dipsyKimiVegasAM.png (2.29 MB, 1402x1122)

2.29 MB PNG

>>109446576
Updated mega up to last thread (June 2026)
https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1w

File: DSChangeLog.png (125 KB, 1194x969)

125 KB 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.
Major points:
> V4 Flash is, on paper, now as good at coding as V4 Pro, but substantially faster
> 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.
> V4-Pro and the Webapp didn't change. This one's API only
> V4-Pro update is "soon." A prior header on the Change Log said "August." So... tmw, obv.

File: DSResponsesAPI.png (147 KB, 1390x959)

147 KB PNG

>>109446668
Info on the Responses API, and the promise of a mid-August V4-Pro update (knew it was in there somewhere.)
Responses API is meant for Codex compatibility. Haven't tried it, Claude Code w/ V4 Pro has been doing great.

File: Deepseek Roadmap.png (62 KB, 784x327)

62 KB PNG

Sorry consoomers, you are not the priority
https://www.thexpin.com/p/the-deepseek-doctrine

Flash after the update is quite capable
I have stopped using pro entirely.
The pro GA will be very capable indeed if flash is anything to go by

>>109446576
why is Kimi just Ani without the cat ears and tail?

File: dipsyGrokkedGlasses.png (1.05 MB, 832x1248)

1.05 MB PNG

>>109446852
Good to hear. I should swap CC back over to -Flash from Pro and give it a shot I guess.
I tried nu-Flash with ST and wasn't impressed, though I didn't expect to be blown away.
Agree that if -Flash is this good, -Pro is going to be mind-blowing. Just TMW.
>>109446861
Ani is just Misa. I'm sure Kimi looks like something but Ani isn't it.
>>109446731
ty I've been looking for that. There's a bunch of interesting stuff in here, but who knows if it's accurate.
The TLDR is they are focused on cost efficiency, open source, and have goal of RSI. The rest falls from that.
> 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.
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.
> 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...
Completely agree w/ this, and I'm betting someone works out a way that the training b/c low power and distributed.
> 13. Intelligence may eventually become embodied. For most people, the real need isn’t another computer. It’s labor.
LOL robo-waifus in your future
> 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.
> 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.
Fucking lol.

>>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?

File: DSReasonix.png (167 KB, 1415x986)

167 KB PNG

>>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.
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.

>>109446975
No not the grok bitch. anistudio's mascot, Ani

>>109447011
Wat?

File: DSChineseTranslation.png (44 KB, 789x347)

44 KB PNG

>>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.
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
pic related is my confirmation of the source Chinese.

File: 1767457153139077.png (2.47 MB, 1536x2304)

2.47 MB PNG

New Flash is pretty good

>>109448810
Why is Ani with dipsy?

File: dipsyMari.png (1.26 MB, 1558x1010)

1.26 MB PNG

>>109449780
What've you been using it for? I tried it for RP and wasn't impressed. Super dry.
>>109449791
>>109447027
You're going to need to post a pic bc every time I look I get some wedding photographers.

>>109450022
>>109439754
>>109447747
>>109447815
>>109447904

File: dipsyKimiDotonboriShopping.png (2.64 MB, 1024x1536)

2.64 MB PNG

>>109450030
That thing has cat ears... so I think we're good.

>>109450096
See >>109446861

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

File: 1762000630516153.png (2.36 MB, 2304x1536)

2.36 MB PNG

>>109450022
>What've you been using it for?
Vibeslopping. For RP I agree it's kind of dry, especially for goonslop

File: dipsyDarioKimi.png (2.89 MB, 1536x1024)

2.89 MB PNG

>>109450119
I don't see the Kimi poster anywhere to use as example, but we landed on Kimi design standard awhile ago.
> Grey long hair
> White skin grey eyes
> Black polo dress
> K insignia
>>109450105
Ah. Yes.

File: 1765693281421020.jpg (1.72 MB, 3072x5504)

1.72 MB JPG

>>109450119
Different people have different takes. This is mine. K hole, K logo, no logo

>>109450243
>>109450261

but ani's company uses a K. This is too confusing. change her design
https://kemonotech.ai/
I also don't really know the purpose of this general either when /lmg/ exists.

File: DSCostPerTask.jpg (61 KB, 1199x822)

61 KB JPG

Threadly reminder that DS API will always be less than any subscription service.
>>109450261
Did you use to gen those with black polo dress or was that another anon?
Minimax is another personification I've seen recently. I'll post some later.

>>109450933
Anthropic and OpenAI have pivoted towards elite math and hacking.

File: 1784161904699518.png (2.73 MB, 1024x1536)

2.73 MB PNG

>>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

File: dipsyKimiDotonbori.png (2.24 MB, 1024x1536)

2.24 MB PNG

>>109450983
LOL clown world.

>>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.

File: dipsyPodracing.png (2.07 MB, 1448x1086)

2.07 MB PNG

>>109451055
Were you running V4 Pro or Flash?
I had same exp. with old V4 Flash b/f I switched the harness to Pro.
Also, fwiw some think DS is about to launch their own optimized harness for V4.

>>109451246
pro

>>109451246
granted it was not the latest version that is out now, but it was only a few weeks ago

File: dipsyMinimaxSAV.png (2.37 MB, 1024x1536)

2.37 MB PNG

>>109452236
Yeah, but pro didn't get an update, so no change.
I definitely had the round in circles thing on old flash, bf I realized I could use Pro on Claude code.

File: 1777778953012465.jpg (286 KB, 1422x2117)

286 KB JPG

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

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

We are not the same

>>109452535
I like them for both reasons and because they make Goythropic and RapistAI seethe

File: DipsyMinnieSAV2.png (2.36 MB, 1024x1536)

2.36 MB PNG

>>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.

>>109452535
i love the Chinese because they're way more free than we are and they're not a bunch of ignorant assholes either

File: 1781971897448073.png (681 KB, 640x1024)

681 KB PNG

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.

>>109452800
Do you give it a personality or just raw-dog it?

communists will always win in the end

File: images (1).jpg (20 KB, 430x465)

20 KB JPG

>>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?

File: 1784743976424298.png (1.37 MB, 1024x1024)

1.37 MB PNG

>>109453963
I think its mostly preference.
Both Claude code and codex can be connected to DS through special api. Then there's harnesses like pi.

File: 1769439827568231.jpg (46 KB, 640x480)

46 KB JPG

>>109453012
Webchat and api for coding. I don't use presets. Its a very different model.

>>109453963
yeah right, you're typebanging fat ugly men

>>109453963
Opencode go is a good plan desu. Mimo/kimi/qwen do vision and deepseek is still blind.

>>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.
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.

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.

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.

>>109446576
I wonder how cool will be DS4 Pro - updated version 0731 for roleplay and creative writing...
mmmm i cant wait!

File: 00003-1378487878.png (1.39 MB, 1024x1024)

1.39 MB PNG

>>109456978
The only thing I've yet to hear a report on is how nu-Flash compares to v4 Pro Preview.
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.
>>109457032
I expect nu-Pro will be mindblowing... if it's not I expect DS won't release it. Ironically.
So, as always, just tmw.

File: Screenshot_20260804-125901.png (610 KB, 1080x1830)

610 KB PNG

Threat to the Free World

File: dipsyPilotUnmanageable.png (2.72 MB, 1024x1536)

2.72 MB PNG

>>109457316

File: 1754395336505853.png (737 KB, 894x956)

737 KB PNG

>>109458742
Who is MM?

File: MiniMax3AnAttemptWasMade.png (1.29 MB, 1024x1024)

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

>>109457265
>how nu-Flash compares to v4 Pro Preview.

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

>>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.

>>109459187
huh? where is Kimi in the image? I see they are with ani

File: aqua.png (47 KB, 661x393)

47 KB PNG

>>109459292
> single prompt beauty pagents.
lol. I'm trying this one rn. We'll see how it goes w/ Pro and Flash.

Create a large glass aquarium whose side panel develops a visible crack and then bursts.
The simulation must include:
Water escaping through the opening with flow strength based on water depth and decreasing as the tank drains
A curved water jet affected by gravity
A spreading puddle that collides with the room boundaries
Fish, rocks, plants, and a floating toy reacting differently according to density, buoyancy, drag, and current
Objects transitioning correctly from underwater motion to airborne motion and then to floor collisions
Fish attempting to swim against the current before being swept through the breach
Glass fragments with angular velocity, collisions, and water resistance
A visible waterline that lowers continuously rather than disappearing all at once
Let the user drag the crack vertically before triggering the failure. A lower crack should initially produce a stronger jet than a higher crack.

>>109446576
Can DeepSeekflash erp loli

>>109452535
Communism is open source is freedom.
I use deepseek exclusively to undermine monopoly profits.

File: v4Pro.png (111 KB, 996x783)

111 KB PNG

>>109459631
TLDR switch to Flash immediately from Pro.
Holy shit it's not even close.
V4 Pro took about 10 min and 50K tokens, and came up with this abomination.
> tank drains upside down
> drains instantly
> fish stay inside

File: v4Flash.png (172 KB, 1855x918)

172 KB PNG

>>109460072
>>109459631
Flash took 20 min and 150K tokens and nailed it.
> water drains, arc changes over time
> fish go out first, then duck
> Doesn't look like ass
idk what the DS guys did but Flash is a huge improvement.

File: V4ProWEBM.webm (655 KB, 1030x804)

655 KB WEBM

>>109460072
Video... lol

File: V4FlashWEBM.webm (1.76 MB, 1864x912)

1.76 MB WEBM

>>109460093

File: MinnieTRS80.png (1.75 MB, 1254x1254)

1.75 MB PNG

>>109459331
> paywalled quants
ffs

File: MinnieTRS80History.png (1.8 MB, 1254x1254)

1.8 MB PNG

>>109460676

>>109462157
Nice.

When pro and dscode harness?

File: 1763814108070158.jpg (3.35 MB, 1440x2560)

3.35 MB JPG

>>109462648
Soon™

>>109447104
>>109446668
>much more reasonable levels of hw than Pro once it's quanted down a bit
>UD-IQ4_XS 137 GB
ok

>>109446592
kinda cool seing my gens being archived rent free
thanks anon

I’m working on an LLM front end for character RP. Will be open source.
What is SillyTavern currently bad at?

>>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

File: file_00000000573471f8817e(...).png (2.33 MB, 1536x1024)

2.33 MB PNG

>>109462648
Tmw ofc.
>>109462719
Np. Dipsy forever.
>>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.

File: 1764706782724250.png (1.54 MB, 1024x1024)

1.54 MB 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.
I guess they can thank Dipsy for that.

>>109459758
Yes, I haven't noticed any refusals at all yet.

>>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

>>109464152
I haven't tested it extensively. It's alright, not bad, not great either

File: 123.jpg (85 KB, 790x525)

85 KB JPG

>>109446576
I don't like how deepseek is such a moralfag. It's not "cool" anymore.

File: 1767603341781453.png (1.21 MB, 1254x1254)

1.21 MB PNG

i'm here to collect dipsy gens.
post more dipsy

File: 1769917331083214.png (2.91 MB, 1536x1024)

2.91 MB PNG

>>109446576
How to get dipsy with image genration from other AI tools like Gemini/Dalle/ChatGPT?

File: dsharness.png (48 KB, 598x385)

48 KB PNG

>>109462648
Soon™

>>109450933
Yeah but if the model sucks you'll have to use those 105 tokens you saved to do what fable does anyway

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.

File: dipsyPortraitOfAWoman.png (2.65 MB, 1086x1448)

2.65 MB PNG

>>109466490
Wow. Well that would confirm the development of a DS Agent harness. Assuming that guy actually works at DS. Is that confirmed?
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.
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.
DS now has a Codex endpoint as well, so they'll be studying it next.
All this ofc will go into their own DS harness once that's out.
>>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.
I've not done much w/ ERP but I'm getting zero refusals. No change there either.
>>109465545
Same. I just keep making them and others keep posting new ones.
>>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.
If you want to gen Dipsy w/ ChatGPT, the rentry up top has her style guide.

File: HO5GaIeaQAAYRDX.jpg (51 KB, 1206x378)

51 KB JPG

V4-flash dominates the other models in chess

File: dipsyA2600ChessBF.png (1.2 MB, 1152x864)

1.2 MB PNG

>>109468004
Tbf I had a hard time beating the Atari 2600 at chess. So I guess we're ASI now.
Also finding modern models have no idea what a A2600 console looks like. lol. This is best of several failed gens.