DeepSeek’s new model cuts API costs in half

Rizzle

Watch DeepSeek’s new model cuts API costs in half - Rizzle on Dailymotion

Transcript

00:00DeepSeq's New Model Cuts API Costs in Half

00:03DeepSeq, the low-key Chinese AI lab that loves surprising the industry, popped back into the spotlight with a brand new experiment.

00:11On Monday, the company quietly dropped an experimental model called V3.2XB on Hugging Face, along with an academic paper on GitHub.

00:21The pitch, slash the price of running Big AI, models on really long conversations or documents.

00:27DeepSeq calls it sparse attention, which is a neat bit of engineering.

00:31The system first deploys a lightning indexer to cherry-pick the most important chunks,

00:35then a fine-grained token selection system grabs only the keywords or tokens.

00:40The result, the model pays attention where it matters and ignores the fluff,

00:44like an over-caffeinated editor skimming a 600-page novel for plot twists.

00:49Why does this matter?

00:50Inference is expensive.

00:52It's the daily grind of serving billions of user queries, not training, that kills your wallet.

00:57DeepSeq claims that for long-context tasks, its method can cut API costs by half,

01:02and the weights are open and free for Hugging Face Tinkerers.

01:06This isn't DeepSeq's first rodeo.

01:08Earlier this year, the company grabbed headlines with R1,

01:11a reinforcement learning model that promised a cheaper path to cutting-edge AI.

01:16R1 didn't exactly spark the revolution some predicted,

01:19and DeepSeq slipped back into stealth mode until now.

01:23V3.2 XP probably won't break the internet the way ChatGPT did,

01:27but its thrifty new attention system could nudge the entire industry toward leaner, meaner AI.

01:33In a world where every extra token costs money,

01:36that's a story worth paying attention to, even sparsely.

01:39www.youtube.com slash.com so forth.

01:40www.youtube.com hashtag.com

01:40something that wants to do is use beria file,

01:42companies choose to use with the user

01:52or the transportation system can run the amount of data if you want to hide.

01:57But the people with the U.S are looking at, we also need to see that you rarely see thefuller ideas and see if you know the technical conditions.

Category

Transcript

Be the first to comment

Recommended