Hacker News Reader

You won't need a frontier size model for most tasks before long. Qwen 3.6 (small) punches way above its weight. I run it at home @8bit on an OEM Spark

View on HN

Second this, I am also running qwen 3.6 35b Q8 on a 5090 liquid getting around 250 tokens / second and it is plenty capable. I actually haven't even looked at models recently because I am happy with what I have.

And.. now I feel the need to look again. Darn, there goes my afternoon

And corporations could run DeepSeek models on cloud hardware.

You can run most open models on cloud hardware. Google Cloud gives you a click to deploy, but then you have saturation / ROI considerations, versus Google serving them up multi-tenant, per-token.

For ROI though you can run 24/7 agentic-style workloads, constantly churning through all your source code looking for security bugs (or whatever) and you DONT pay per-token costs.

A DeepSeek instance running 24/7 in a cloud provider will beat doing that with Claude which could bankrupt you with 100x more costs, even though it might find more.

And DeepSeek may find enough to keep your engineering team saturated and busy fixing things.

This process works better with multiple models and not simply slinging Ai at it 24/7. It's not an ROI just because you keep the GPU busy. The signal to noise ratio is not there yet

Reply on news.ycombinator.com