Deepseek has released DeepEP V2 and TileKernels.
Posted by External_Mood4719@reddit | LocalLLaMA | View on Reddit | 18 comments
Posted by External_Mood4719@reddit | LocalLLaMA | View on Reddit | 18 comments
SilentDanni@reddit
I’ve no clue what these folks are up to but it occurred to me that they’re doing what OpenAI was supposed to do. They’re actively advancing research and sharing their findings. I’m sure they have a lot of proprietary technology but it builds some good will when a company is willing to open source some of the stuff they’re working on. I’ll be digging into it tonight.
PoopSick25@reddit
Deepseek was never a commercial product company. They are first and foremost research lab and we have a lot to thank their research for how the ai models are how they are rn
Beginning-Window-115@reddit
they use it for trading
Fantastic-Balance454@reddit
That's oversimplification, sure the team that started doing this was working at a quant. But at this point they've certainly pivoted beyond that.
TamSchnow@reddit
I saw a funny sketch once about this.
„Why did you release the model for free??!?“
„So that I could short NVIDIA stock“
Mochila-Mochila@reddit
Dayum, it's both diabolical business-wise, and altruistic public-wise 🤯
sk1kn1ght@reddit
Yeap. I read the merge and the repo. It's English words alright!
Pyros-SD-Models@reddit
You should look up how many papers OpenAI is realeasing in a year and you will end up with a number higher than deepseek papers but OpenAI bad obviously. Needs to be mentioned even in non-OpenAI related threads lol.
power97992@reddit
I hope this is alluding to the release of v4…
DefNattyBoii@reddit
Remeber, V4 is always a couple of weeks away
FullOf_Bad_Ideas@reddit
Speculation but SM100 is Blackwell that Deepseek is supposed to not have. They could have just rented out 8x B200 on Vast tho technically, but if they're making software for it, they probably are using Blackwell GPUs for training.
Operation_Ivy@reddit
I thought they said no Engram for V4
FullOf_Bad_Ideas@reddit
I didn't see that. Source?
Operation_Ivy@reddit
Not confirmed either way, but see https://x.com/i/status/2047191920888139778 for example. I can't find it but there was a (usually) reliable source that claimed no Engram in V4
TheDeviceHBModified@reddit
Sad if true. That was supposed to be the biggest game-changer in the new generation.
AlwaysLateToThaParty@reddit
Am I reading it right, that they have figured out a parallelization technique that scales linearly? As in; if you double capacity, you double speed?
Saltwater_Fish@reddit
Nice
JobAsleep6653@reddit
DeepSeek commands respect! No matter what they do.