News for hackers. Stuff that matters.

A comment on DeepSeek-R1 with Dynamic 1.58-bit Quantization. Parent comment

Re:DeepSeek-R1 with Dynamic 1.58-bit Quantization by sureglymop on Tuesday January 28, 2025 @07:46PM (#42860075)
I have Kagi Ultimate and it is nice for this. But a cheaper suggestion would be to use OpenRouter and then use these models via Fireworks or TogetherAI. It also integrates into much more applications. AFAIK Kagi doesn't document a user facing API for the assistant feature.
[ Parent | on HN ]
Re:DeepSeek-R1 with Dynamic 1.58-bit Quantization by phaedrix on Tuesday January 28, 2025 @10:46PM (#42861290)
Unfortunately those are both 10-15x the cost of deepseek direct.

Deepinfra is pretty cheap though as a deepseek provider.

[ Parent | on HN ]
Re:DeepSeek-R1 with Dynamic 1.58-bit Quantization by pants2 on Wednesday January 29, 2025 @01:46AM (#42862254)
That's because DeepSeek is subsidizing their API massively to get more training data.
[ Parent | on HN ]
Re:DeepSeek-R1 with Dynamic 1.58-bit Quantization by sureglymop on Wednesday January 29, 2025 @11:28AM (#42867103)
Sure. I meant moreso that this would be cheaper than Kagi while providing the same selection of models.

As for deepseek, I couldn't even sign up because my email domain is not on their whitelist. To just try it out for now I don't mind the increased cost.

[ Parent | on HN ]