Cerebras has put Qwen 3.8 27B on its inference service at 1,500 tokens per second, according to the company's own model documentation. Commenter gpugreg tried it on a coding task and hit the platform's token limit in about 90 seconds, spending $1.10 in the process. Running the same task on DeepSeek-V4-Flash finished in 172 seconds for $0.024, with a final context window of 55,217 tokens. Qwen 3.8 27B, by contrast, was still not done at a 64,178 token context window when gpugreg's run ended.

Part of the problem is what counts against the limit. gpugreg says cached tokens count toward the per-minute cap, which is what burns through it so fast. Commenter nostrebored gives a different number for the same cap, 150,000 tokens per minute on the public endpoint, and says that's too low for many coding tasks. Cerebras did not explain the discrepancy in the source material.

On quality, the comments split. gardnr, who used Cerebras' coding plan for months, calls Qwen 3.8 27B "likely one of the strongest models they've hosted so far" and says the speed changes how it feels to use. ecshafer, running the same model self-hosted, says the opposite: used agentically, it "will spin around in circles on even small tasks talking to itself until it loses context and starts again," and at one point told the user it had forgotten their original question.

gardnr also asks whether Cerebras has added prompt caching yet, saying its absence made agentic coding expensive in past use. Nobody in the thread answers. Separately, nostrebored says their account was switched to an enterprise tier that blocks self-serve billing, with the message "Self-serve billing is not available on Enterprise accounts. Please contact your team for further questions," despite having no team on the account after being removed from Cerebras' Slack channel.

Whether prompt caching arrives, and at what price, is the question gardnr's comment leaves open. Until it does, the fastest model on the board may also be the one that costs the most to actually finish a job.