Artificial Intelligence thread

[9dashline]

Please, Log in or Register to view URLs content!

Looks like deepseek dropping something today

 

[9dashline]

looks like it's best at coding, just under o1 high pro mode

It beats all closed source nonreasoning models

 

[Overbom]

Huge. On LiveBench, DeepSeek V3, is second best non-reasoning model (after gemini-exp-1206). And best open source model

Can't wait to see the reasoning variant of this.

 

[tphuang]

it's actually illogical to continue to put foundational models together with reasoning models. You can't get high speed responses with reasoning models. Having good foundational model is really important.

The fact they can get such a great foundational model and put it on open source is a big deal. Since it's so much better than Llama 3.1.

I wish Alibaba team can put out a higher parameter version of Qwen for public use.

 

[9dashline]

Hopefully before 2024 is over they release deepseek r1 weights, and we get a surprise of only 16b

Now that we have open source better than sonnet 3.5 it will put pressure on claude anthropic to up the ante.