Cerebras adds Qwen 3.8 27B public API endpoint rated at 1,500 tokens/s

I used Cerebras' Coding Plan for a few months, and it is genuinely hard to keep up with the model speed. Qwen 3.8 27B may be one of the strongest models they have hosted, but agentic coding got expensive without prompt caching.

I can see this speed pushing SaaS UX to another level.

It looks like Cerebras supports prompt caching in its inference docs.

I thought they used to have a monthly plan. What happened to it?

Why host smaller models instead of the 2.4T version? I wonder whether wafer I/O and interconnect limits make that impractical.

I assume it is mostly economics. The inference service is also advertising for Cerebras' hardware business.

The wafer has only 44 GB of SRAM. Once they offload memory, they lose the speed advantage of keeping the model on one chip.

I used their Code product with GLM 4.7. It is fun, but a weak model does not get much useful work done. I hope they add stronger models to Code.

I wish customer support were somewhere other than Discord. Discord flags me as a bot and will not accept my email or phone number.

I hit an onboarding redirect loop when creating an account. Google sign-in worked for me.

Qwen 3.8 is not on OpenRouter yet. Cerebras only allows 128k context for it, so it seems better as a specialized sub-agent than for long tasks.

I hope Cerebras exposes this capacity through OpenRouter. The fastest Qwen 3.8 provider there is around 80 tokens/s, while Cerebras lists 1,500 tokens/s.