Qwen 3.8 27B on Cerebras: 1,500 Tokens/s

Qwen 3.8 27B just landed in Cerebras' public model catalog at ~1,500 tokens/s — wafer-scale speed for a 27B open-weight model, with 128K context on the paid tier. It's been on Hacker News for nine hours and is already at 473 points. The fastest OpenRouter provider serving the same weights does about 80 tok/s. That's a 19x gap for identical parameters.

What shipped

The catalog entry (qwen-3.8-27b, 27B params, 64k/128k context, ~1500 tok/s) sits beside gpt-oss-120b at ~3000 tok/s, available on free trial and pay-as-you-go tiers. Early reports from the HN thread: one dev hit the 450,000 tokens/min rate limit in ~90 seconds, burning $1.10 on cached tokens; another clocked a "super fast" output phase but saw tool-calling failures more often than with DeepSeek 4. For local reference, 200 tok/s on an RTX 5090 via ninfer is what "fast" meant 48 hours ago.

Why it matters

This is the pattern to watch: the same open weights, 19x apart in serving speed. At 1,500 tok/s, agentic loops stop being "wait for the model" and become "wait for the rate limit." The bottleneck has moved from silicon to token budgets and tool-call reliability — and those are software problems. For anyone building on open-weight models, Cerebras just redrew the cost-per-token-time curve, and the pricing is aggressive enough that even cached-token burns ($1.10 in 90s) are cheap compared to waiting.