Deepseek V3 is really this good or just hype?

I spent the whole day yesterday testing DeepSeek, solving coding problems with Open Hands (previously called Open Devin).

The model feels rock solid. At times, it would go off track, but a simple reset of the window was enough to bring it back in line. Once that was done, it was smooth again.

DeepSeek has really raised the bar. What do you all think?

Yeah, DeepSeek really got me excited again about AI. The model is smart, and the cost makes it easy to add AI into projects without worrying about the budget. I had an idea for an AI-driven video game when ChatGPT first came out, and now it finally seems possible.

@Rafe
You mean cheap APIs? Because with 685B parameters, not many people can run it locally.

Kip said:
@Rafe
You mean cheap APIs? Because with 685B parameters, not many people can run it locally.

Yeah, APIs. I haven’t compared prices much, but I tried DeepSeek through OpenRouter, and it was fast, smart, and super cheap. I used it for a while and only spent 5 cents on compute.

@Rafe
Wait, you ran a high-end model for a long time and only spent 5 cents? That’s impressive. How did you do that?

Chloe said:
@Rafe
Wait, you ran a high-end model for a long time and only spent 5 cents? That’s impressive. How did you do that?

@Raven
Got it, thanks.

@Raven
Is it better to use OpenRouter or go straight to DeepSeek?

Reagan said:
@Raven
Is it better to use OpenRouter or go straight to DeepSeek?

Not sure if it’s better, but having credit on OpenRouter lets you switch between multiple models without needing to host them or pay separately.

@Raven
Honestly, it depends on how you use it. If you like flexibility, OpenRouter might be the way to go. If you just want DeepSeek, going direct could be easier.

I find it weaker than Claude, but I don’t use it for coding. Surprised it’s getting so much hype.

I just chat with AI about different topics. I’ve tried 4o, Sonnet 3.5, all Gemini versions, Grok, and a bunch of open-source models. DeepSeek is better than most open-source ones, but I don’t think it’s on par with Sonnet or 4o.

It gets stuck in loops sometimes, ignores my prompts, and gives weird responses. Maybe it’s optimized for coding? I’ve used both the DeepSeek chat interface and OpenRouter.

@Sterling
That could be it. I haven’t used it for regular conversation, only for coding.

I actually like using the real-time Gemini API for chat.

Zev said:
@Sterling
That could be it. I haven’t used it for regular conversation, only for coding.

I actually like using the real-time Gemini API for chat.

Same. I use the multimodal Gemini API more than ChatGPT’s voice mode. The only issue is the 15-minute limit. Gemini 2.0 follows instructions better than most models, especially for roleplay.

Is it cheap to run locally?

Oli said:
Is it cheap to run locally?

Not at all. It’s a massive model. The API pricing is surprisingly good, though.

Noor said:

Oli said:
Is it cheap to run locally?

Not at all. It’s a massive model. The API pricing is surprisingly good, though.

Yeah, but it’s on a discount right now. The price goes up after February.

@Oli
Even at full price, running it locally is still way more expensive than using the API.

Noor said:
@Oli
Even at full price, running it locally is still way more expensive than using the API.

It’s a mixture of a big model and a MoE (Mixture of Experts) model. It activates about 37B parameters per response. Running that on a CPU is possible but painfully slow.

Is there any provider hosting this model in North America? I don’t want my data going to a Chinese server.

Ainsley said:
Is there any provider hosting this model in North America? I don’t want my data going to a Chinese server.

Yes, https://fireworks.ai/ is an American company hosting it.