Cerebras runs Qwen 3.8 27B at 1,500 tok/s — caching costs full price
The fastest hosted open Qwen yet at $0.99/1M in — but every cached token bills at full rate, and the same weights run free on a Mac.
Copy markdown
1,500 tokens/second, no waiting
Cerebras now serves Qwen 3.8 27B at roughly 1,500 tokens/second on its pay-as-you-go tier — about 2.8x faster than typical GPU providers. For agentic loops and long reasoning chains, that's the difference between watching a spinner and getting an answer.
The catch: cached input bills at full rate
Cerebras gives no discount on cached or repeated tokens, so prompt-heavy agent runs pay the full input rate every turn. One tester's 5-minute coding session cost $1.60 versus $0.29 elsewhere — roughly 5.6x more, despite being ~2.8x faster.
128K context and a 150K/min ceiling
The endpoint caps context at 128K (below Qwen's native window) and throttles at 150K tokens per minute, cached tokens included. List price is $0.99/1M input and $1.49/1M output, and Cerebras itself flags the shared tier as not for production.
It quietly replaced Gemma 4 31B
Qwen 3.8 27B took over the same Cerebras slot Gemma 4 31B held, with no migration notice. If your code still points at that endpoint, the model underneath it changed — re-test your prompts before you trust the output.
The free alternative: run it on your Mac
The same 27B weights run locally on Apple Silicon — a 4-bit build fits in about 15GB and pushes 100+ tokens/second on high-end M-series chips. You trade Cerebras's raw speed for zero per-token cost and the full context window.