What's this about?
Cerebras is serving Qwen 3.8 27B at 1,500 tokens per second — an order of magnitude faster than typical GPU serving (the fastest OpenRouter provider does ~80) — with prompt caching supported but not discounted.
What exactly is Cerebras?
A chipmaker whose processor is one dinner-plate-sized wafer instead of many small GPUs — the whole model sits in on-chip memory, eliminating the memory-shuffling that bottlenecks GPUs. Blazing generation speed is the payoff; capacity and rate limits are the cost.
What's the community saying?
HN 467 points, 134 comments
- The rate-limit wall: 150–450K tokens/minute caps make it hard to actually use for agentic coding — one user hit the limit in 90 seconds and burned $1.10, because *cached* tokens count against the limit.
- Hands-on nuance: output is stunning, but input processing isn't faster, and tool-calling fails more than competing models — retries eat the speed win.
- Why only small models? — a genuine architecture debate about wafer memory and interconnect limits.
- A quality split: some find 27B one of the strongest models Cerebras has hosted; a self-hosting user finds the same model "unusably bad" agentically. Harness and serving stack, again.