darkevilmac

DeepSeek models are mainly focused on the efficiency they manage to have and how it forces other labs to push their prices down.

I've been able to use V4 Flash for most things over the past two weeks without really feeling like I need to use more expensive options like Kimi K3 or any of the American labs.

I'm mainly referring to their cloud offering if someone needs European hosted models

I don't really pay attention beyond that given that basically every company in this industry is doing something stupid

Well, this one will be a little bit, but Ollama hosts in Europe and has options for no data retention

Micron

Does not work on Voyager for Android, the humble inefficient GIF continues in spite of more efficient options

Counterpoint: GIFs loop by default in basically every app, WEBM doesn't

I like this article downplaying the model by saying this is only a "mini deepseek moment in a developer centric sense" - as if developers aren't the main profitable use case for LLMs and the ones responsible for the model selection in products that reach end users

If the model is good enough for a developer and the cost is right, then that's the model that you're going to be using whether you notice or not.

I've used local models and they just tend to screw up more often in my experience. But I'm also more focused on having agents do long running tasks which small models just aren't good at.

Well then Alibaba needs to get better at it cause the Qwen models have kinda sucked in my experience.

Well your first mistake was using Chrome in 2026