@antirez Thank you!

@antirez Ok magic man

@antirez Unboxed my m5 and downloading weights right. Can’t wait to get started tomorrow. Thanks antirez!

@antirez “AI is too critical to be just a provided service»

Thank you for your great work 💪

@antirez It'd be interesting to look at the optimal coding agent shape given the local AI constraints: <1M context, fast prefill, slow-ish decode, and free/infinite tokens.

@antirez Having spent the whole day benchmarking models for my case (follow instructions with limited freedom), DeepSeek certainly isn’t “quasi-frontier” for me anymore.

It’s on-par with Gemini and both came out ahead of GPT and Opus. It truly is an amazing model!

@antirez I can't sing enough praises about dsv4-flash either. It is an insanely high quality model.

@antirez ds4.c is insanely good for the local models scene. It's almost as good as llamacpp was. I bet you started a new trend for building model--specific inference engines for exact hardware setups, i.e. 128gb unified memory (no matter is it strix halo, or spark or m-chips). That huge.

@antirez 🙇🏻‍♂️

@antirez Thank you for doing this!

@antirez ds4 is exactly the kind of inference engine that makes the .84/month number possible. Running DeepSeek models efficiently on consumer hardware is the unlock. Your work on this is the infrastructure layer most people don't appreciate.

@antirez Nice read, @deepseek_ai Thank you for flash model.

@antirez already have 5 tabs open and now this

@antirez Dear antirez, just a random question: now that AI coding agents have become so capable, and a lot of developers are relying less on writing code entirely by hand, I’m really curious ,what’s your take on this shift?

@antirez Congrats and thanks for creating one more amazing project. Grazie!

@antirez Have been using DS4 for last couple of days on dgx spark. Truly impressive. Thank you !