Cerebras support moving to "Best Effort"

[acoliver]

I was really excited when Cerebras offered their subscription but after a few months of this I've cancelled. The value is rather poor and their hoist of each model evals more poorly. Meaning if you're purely vibe coding "do this" turn "okay now do this" then you might value the speed advantage (which is much lower than they claim even according to their own endpoint). However, if you do any automation at all you'll realize it takes more passes for Cerebras to get things right because the model is brained just a little and then there is less context to boot. So ultimately you're better off with a slower but better provider.

I've been using Synthetic.new a lot. Like most of them, it isn't entirely stable but the value is good and they are open and honest when things go wrong. Also having a bit more selection is nice (MiniMax, GLM 4.6 and K2-thinking -- other models are proxied to fireworks and a bit slow). It won't be that we don't support Cerebras in LLxprt -- but their endpoints are a bit weird. We just aren't running out Github Actions against it anymore (moved to Synthetic) and aren't using it to develop LLpxrt. That means that we'll be more dependent on user reports to fix issues.

Just wanted to make that clear. If you're looking for providers to use with LLxprt I wrote a bit about it in InfoWorld