Deepseek casually drops a 685 Billion Parameter model on Hugging Face with MIT license! The higher variant performance on par with GPT-5 and Gemini 3.0 Pro! And yet, this model is open weights coming with a Technical Paper!
The new Deepseek is here! Gold-medal performance in the 2025 International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI).
I prefer my privacy invading apps to be watched by the US intelligence community.
It's an open weights model. You can download it and run it if you have a big enough computer. It's like pirating movies and watching on your computer. No one will know from the movie studio. Except here you don't have to pirate, it's free to download and redistribution.
These behemoth parameter models are great, especially when hosted externally in the cloud. Good instructional models that run locally with sub 16 billion parameters with solid tool access and live internet access will win the day, at least in my book.
I guess they or someone using their models will release the distill versions
Bored of the same old CS skins? Time for an upgrade! Unlock rare CS2 drops, swap your favorites, and hit the match with a fresh new look. Make a deposit now and claim your BONUS!
Wow, LittleCoder, that's a game changer, but open source might face challenges, no?
Its defaulting to response in Chinese to me. My language setting is system
Open source with this super large LLM are good as unusable for any normal person. How can can you run this locally without spending 10s of thousands in GPU? It's just a marketing ploy really.
It's great we have open source like Deepseek keeping the paid models more honest. Can't wait until they get good enough to be run on the average person's PC.
great for the opensource community, and some important proof of their new attention working but this type of performance is almost standard with the public recipe of making models. The bigger question is why aren’t the frontier labs further ahead, maybe they only release models a…
Want to reach a similar performance? 📅 Register to our Webinar 'X Ads for Beginners' to learn how!
DeepSeek dropping this beast on Hugging Face feels like the open source revolution just leveled up big time and with its sparse attention slashing inference costs by 70% that MIT licensed powerhouse turns elite reasoning into everyday reality for devs everywhere.
V3 is actually a 671B MoE with about 37B active params, trained on 14.8T tokens in roughly 2.8M H800 hours, using MLA and multi token prediction. Frontier level performance with open weights and MIT license is a crazy unlock for anyone building serious agents
Is this really on GPT-5 level or is that just benchmark talk? Anyone tried it yet?
China, casually wiping out moats. They seemingly want to fill the world with machine intelligence that runs on everything. Just don't ask it about Tianmen square 😁 The PRC being more Open AI than OpenAI wasn't in my futures checklist.
This is how it should be. No hype, no drama, pure engineering. All the other releases are nothing but more drama less impact.
still common man can not download and use it. Need high end costly system. Even Quantisation wil be higher
in reality it comes as close as like 30% maybe not even that to models like gemini 3 or gpt5. Thats the reality, these models are trained for testing, not for actual usage. They are useless in practice
685B open weights with MIT license is wild. The open frontier just leveled up.
China is on a level the West cannot comprehend. Elon, Dalio, and Sachs understand fully, world’s richest man, world’s largest hedge fund manager, and top 3 diplomats in the world. But what do they know…
When casually means 685 billion parameters and performance that rivals the big players, that's not casual that's revolutionary 🚀
Open weights + MIT license + GPT-5-tier performance… This isn't a release. This is a declaration of war on closed-model economies.
GPT-5 & Gemini 3.0 Pro, watch out! 🤣 Open source is coming for you! Seriously though, congrats to Deepseek on this impressive release. Can't wait to try it out.
United States Trends
- 1. World Cup 200K posts
- 2. Paraguay 21.1K posts
- 3. FINALLY DID IT 426K posts
- 4. The Jupiter 96.7K posts
- 5. Morocco 60.5K posts
- 6. Croatia 14.5K posts
- 7. Argentina 184K posts
- 8. Infantino 54K posts
- 9. Portugal 74K posts
- 10. #USMNT 1,196 posts
- 11. Matt Campbell 9,350 posts
- 12. Wayne Gretzky 3,448 posts
- 13. Group D 14.6K posts
- 14. Senegal 33.6K posts
- 15. Lauryn Hill 10.3K posts
- 16. Iowa State 8,048 posts
- 17. Warner Bros 208K posts
- 18. Norway 27.9K posts
- 19. Ghana 61.2K posts
- 20. #Mundial2026 27.8K posts
Something went wrong.
Something went wrong.