vhstape@lemmy.sdf.org
on 30 May 00:24
nextcollapse
the Chinese AI lab also released a smaller, “distilled” version of its new R1, DeepSeek-R1-0528-Qwen3-8B, that DeepSeek claims beats comparably sized models on certain benchmarks
Most models come in 1B, 7-8B, 12-14B, and 27+B parameter variants. According to the docs, they benchmarked the 8B model using an NVIDIA H20 (96 GB VRAM) and got between 144-1198 tokens/sec. Most consumer GPUs probably aren’t going to be able to keep up with
On my Mac mini running LM Studio, it managed 1702 tokens at 17.19 tok/sec and thought for 1 minute. If accurate, high-performance models were more able to run on consumer hardware, I would use my 3060 as a dedicated inference device
brucethemoose@lemmy.world
on 30 May 18:15
collapse
Depends on the quantization.
7B is small enough to run it in FP8 or a Marlin quant with SGLang/VLLM/TensorRT, so you can probably get very close to the H20 on a 3090 or 4090 (or even a 3060) and you know a little Docker.
TropicalDingdong@lemmy.world
on 30 May 00:43
nextcollapse
Yeah idk. I did some work with deepseek early on. I wasn’t impressed.
HOWEVER…
Some other things they’ve developed like deepsite, holy shit impressive.
Above is what I can do with deepsite by pasting in the first page of your lemmy profile and the prompt:
“This is double_quack, a lemmy user on Lemmy, a new social media platform. Create a cool profile page in a style that they’ll like based on the front page of their lemmy account (pasted in a ctrl + a, ctrl + c, ctrl + v of your profile).”
It not perfect by any stretch of the imagination, but like, its not a bad starting point.
knighthawk0811@lemmy.world
on 30 May 00:57
nextcollapse
it’s distilled so it’s going to be smaller than any non distilled of the same quality
LainTrain@lemmy.dbzer0.com
on 30 May 05:23
collapse
I’m genuinely curious what you do that a 7b model is “trash” to you? Like yeah sure a gippity now tends to beat out a mistral 7b but I’m pretty happy with my mistral most of the time if I ever even need ai at all.
threaded - newest
7b trash model?
Most models come in 1B, 7-8B, 12-14B, and 27+B parameter variants. According to the docs, they benchmarked the 8B model using an NVIDIA H20 (96 GB VRAM) and got between 144-1198 tokens/sec. Most consumer GPUs probably aren’t going to be able to keep up with
It proved sqrt(2) irrational with 40tps on a 3090 here. The 32b R1 did it with 32tps but it thought a lot longer.
On my Mac mini running LM Studio, it managed 1702 tokens at 17.19 tok/sec and thought for 1 minute. If accurate, high-performance models were more able to run on consumer hardware, I would use my 3060 as a dedicated inference device
Depends on the quantization.
7B is small enough to run it in FP8 or a Marlin quant with SGLang/VLLM/TensorRT, so you can probably get very close to the H20 on a 3090 or 4090 (or even a 3060) and you know a little Docker.
Yeah idk. I did some work with deepseek early on. I wasn’t impressed.
HOWEVER…
Some other things they’ve developed like deepsite, holy shit impressive.
Save me the search, please. What’s deepsite?
tmpweb.net/nmS9uRBAENhQ/
Above is what I can do with deepsite by pasting in the first page of your lemmy profile and the prompt:
“This is double_quack, a lemmy user on Lemmy, a new social media platform. Create a cool profile page in a style that they’ll like based on the front page of their lemmy account (pasted in a ctrl + a, ctrl + c, ctrl + v of your profile).”
It not perfect by any stretch of the imagination, but like, its not a bad starting point.
if you want to try it: huggingface.co/spaces/enzostvs/deepsite
Excuse me… what? Ok, that’s something…
Here I’m DM"ing you something. Its very personal, but I want to share it with you and I made it using Deepsite (in part).
Ok
👀
it’s distilled so it’s going to be smaller than any non distilled of the same quality
I’m genuinely curious what you do that a 7b model is “trash” to you? Like yeah sure a gippity now tends to beat out a mistral 7b but I’m pretty happy with my mistral most of the time if I ever even need ai at all.
So can a lot of other models.
“This load can be towed by a single vehicle”
ew probably still censored.
You can self host it right??
if the model is censored… then what, retraining it? Or doing it from scratch like what open-r1 is doing?
The self hosted model has hard coded censored content.
The censorship only exists on the version they host, which is fair enough. If they’re running it themselves in China, they can’t just break the law.
If you run it yourself, the censorship isn’t there.
Yeah, i think the censoring in the LLM data itself would be pretty vulnerable to circumvention.
Untrue, I downloaded the vanilla version and it’s hardcoded in.