正在加载视频...

视频加载失败

DeepSeek v4.1 Flash via API is almost TOO fast 🤯

139,524 次观看 • 2 个月前 •via X (Twitter)

46 条评论

keys 🧪 的头像
keys 🧪2 个月前

Actually don’t want the model to run too fast then you have keep scrolling up to see what happened it’s annoying after a while

Mia 的头像
Mia2 个月前

Good problem to have imo

hkdom 的头像
hkdom2 个月前

You will wow more if you run it with DSH

Mia 的头像
Mia2 个月前

You're probably right

Santanu Sinha 的头像
Santanu Sinha2 个月前

Genuine question.. what harness is that and why is it flickering so much? Also .. yes amazing speed

Mia 的头像
Mia2 个月前

Hermes agent CLI

Santanu Sinha 的头像
Santanu Sinha2 个月前

Ok thanks

kaith 的头像
kaith2 个月前

How’s it’s performance?

Mia 的头像
Mia2 个月前

Looking better than v4 flash

Consigliere 的头像
Consigliere2 个月前

via Cloud API?

Mia 的头像
Mia2 个月前

Yes

Gaurav Bhatia 的头像
Gaurav Bhatia2 个月前

Imagine after thinking for that long it doesn’t complete the work.

Mia 的头像
Mia2 个月前

But it did 😀

Wassollichhier 的头像
Wassollichhier2 个月前

is vision onboard?

Mia 的头像
Mia2 个月前

Not on this one, it's beta

Samuel McHargue 的头像
Samuel McHargue2 个月前

@grok how much is v4.1 flash compared to the older one.

Bear 的头像
Bear2 个月前

Vision support?

Daniel Carneiro 的头像
Daniel Carneiro2 个月前

WTF is that ugly TUI?

nah 的头像
nah2 个月前

Imagine every model being this fast. That'd be amazing.

Steven Cheng 的头像
Steven Cheng2 个月前

Latency that low feels like cheating.

base 的头像
base2 个月前

get in my spark … lol

CV.YH 的头像
CV.YH2 个月前

Wow

Chris Winslow 的头像
Chris Winslow2 个月前

Speed ≠ Quality?

Saurav 的头像
Saurav2 个月前

actually useful?

Mia 的头像
Mia2 个月前

Seems like it's better than v4 flash

bitflipgremlin 的头像
bitflipgremlin2 个月前

this isn't quite mercury-2 speed but it's damn fucking close, and quality leads me to suspect it's a different mechanism. i wonder what kind of bullshit the whale came up with now

Ian Hailey 的头像
Ian Hailey2 个月前

It can never be too fast!

dan0mad 的头像
dan0mad2 个月前

There’s 4.1???!? Since wen?

forreal 的头像
forreal2 个月前

better than glm 5.3 flash? it only matters if it reaches performance of kimi k3 or glm 5.3.

Devin Oldenburg 的头像
Devin Oldenburg2 个月前

~176 tokens/sec

Redmix 的头像
Redmix2 个月前

It's a shame we can't enjoy it for very long.

Mia 的头像
Mia2 个月前

Squeezing what I can !

Redmix 的头像
Redmix2 个月前

✌🏻

Cato Nooka 的头像
Cato Nooka2 个月前

if there is another model break through for this kind of speed, I don't know what to say, not only so fast at 400 TPS, but also more token efficient, the time for each task is shink down to minutes

Yume_X 的头像
Yume_X2 个月前

Yeah I was shocked too when I saw it , makes a good argument for 300 tok/s builds when you see it lol

CryptoYeti 的头像
CryptoYeti2 个月前

@grok wie teuer ist die API von der neuste Model von Deepseek? Und vergleiche sie mit Muse spark corbitor .

Jayden 的头像
Jayden2 个月前

The latency is fun. The bill is the actual benchmark.

Mike 的头像
Mike2 个月前

I really wish Nous would fix the CLI redraw flashing in Hermes. That’s mainly why I stick to the TUI.

remakw 的头像
remakw2 个月前

fastest model i gave ever used switched the trace off because my eyes were hurting lol

rapidTools / Mike 的头像
rapidTools / Mike2 个月前

o_0... Dafuq... I hope we get this open as well. It would speed up local flash model az well I guess. (not to this speed obviously, but making it even faster wpuld be cool)

Tom 的头像
Tom2 个月前

Why I only see v40730? Where is 4.1?

Steven Cheng 的头像
Steven Cheng2 个月前

Almost" is doing a lot of heavy lifting there. Latency spikes still kill the vibe when you are chaining calls.

Reelix 的头像
Reelix2 个月前

Do a speed comparison :p

Cave 的头像
Cave2 个月前

Woo.

Jarno 的头像
Jarno2 个月前

Fast until you hit the rate limit, then it's back to staring at a spinner like every other model.

Akshay Joshi 的头像
Akshay Joshi2 个月前

Thought it’s free and went rampant and burned 96million token from Deekseek platform hopefully 97% cache hit hence 1.30$ only

相关视频