News · 2026-09-03

Cerebras is serving an open 27B model at 1,500 tokens a second, and the free tier caps it exactly

Cerebras is serving Qwen 3.8 27B at about 1,500 output tokens per second on its public endpoints, roughly ten times what the same class of model typically achieves on GPUs. The constraint is not the silicon. Cerebras' own rate-limit page caps free-tier users at 90,000 tokens per minute -- almost exactly the model's raw output rate -- meaning the free tier expires the moment you actually use the speed.

Key facts

• Qwen 3.8 27B runs at approximately 1,500 output tokens per second on Cerebras, with 64k context free and 128k on paid, per the Cerebras model catalog.
• Free tier: 30,000 uncached and 90,000 total tokens per minute, at 5 requests per minute. Paid: 150,000 uncached, 450,000 total, 300 requests per minute, per the rate limits page.
• Cerebras targets 10,000 output tokens per second on medium open models with its CS-5 system in 2027.
• Primary sources: Cerebras model catalog and the company's Hot Chips 2026 deep dive.

The model underneath is genuinely open. Alibaba's Qwen team released Qwen3.8-27B on Hugging Face and ModelScope on August 14, 2026, framing it as the first open release of a model in its top-tier Qwen-Max class. The weights are a 55.6 GB download at full precision. Cerebras states that the models it serves publicly are "the original, unpruned versions," with quantization applied only to stored weights -- a claim worth noting because hosted inference providers have quietly shipped shrunken models before, and the difference shows up in quality long before it shows up in a spec sheet.

Here is the arithmetic that makes the rate limits interesting. At 1,500 tokens per second, the model produces roughly 90,000 tokens in a minute. The free tier's total budget is 90,000 tokens per minute at five requests per minute. Those numbers are not a coincidence; they are a fence built exactly at the edge of the field. You can watch the speed happen once, then wait. On pay-as-you-go, the total budget rises to 450,000 tokens per minute at 300 requests -- five times the model's own output rate, which is the point at which the hardware number stops being a demo and becomes throughput you can build on.

Cerebras' documentation also answers, and partly confirms, the standard cost objection. Prompt caching is automatic on the platform, cuts time-to-first-token, and is explicitly designed for multi-turn and agentic workloads. But the caching documentation says cached tokens still count toward the tokens-per-minute quota and are priced identically to normal input tokens. Caching buys latency and consistency. It does not buy cheapness.

The roadmap is the more consequential signal. In its Hot Chips 2026 writeup, Cerebras says its CS-5 system, targeted for 2027, is designed to roughly double CS-4 and reach up to 10,000 output tokens per second per user on medium open models like Gemma 4 31B and gpt-oss-120b, plus 5,000 per user on frontier-scale models -- while supporting models above 50 trillion parameters interactively. The architectural argument is that the Nexus design splits the system into modular compute backpacks, centralises power, integrates cooling and I/O around the wafer, and uses wafer-scale locality to cut the communication overhead that dominates multi-GPU scale-up. Why that matters is covered in our explainer on why LLM inference is memory-bound: at generation time the bottleneck is usually moving weights, not multiplying them, and a wafer-sized chip changes the distance those weights have to travel.

NVIDIA is not conceding the comparison. Its own materials report Gemma 4 31B running on the Vera Rubin LPX platform at 3,400 output tokens per second at 100,000-token context. So the Cerebras critique is narrower than "NVIDIA has no numbers": it is the architectural claim that a dense 31-billion-parameter long-context benchmark does not prove a memory-light decode stack scales into frontier-size workloads where the memory wall is the whole problem.

The skeptics on Hacker News -- 464 points and 275 comments at the time of writing -- argued economics rather than physics. Cerebras is sold out. It sells to enterprise hardware buyers, not individual developers. And one commenter made the sharpest point of the thread: even 1,000-plus tokens per second is worthless if the agent loops on tool calls and burns roughly $5 a minute doing it. Speed multiplies whatever the agent is doing, including the wrong thing.

Ground Truth has covered this platform before, when OpenAI put a frontier model on Cerebras chips at 750 tokens per second. The trajectory is consistent and steep.

The honest caveat: every performance number here comes from vendor materials, on both sides. Cerebras publishes Cerebras' numbers; NVIDIA publishes NVIDIA's. Neither has been independently reproduced on identical prompts, identical context lengths and identical quantization, which is the only comparison that would settle anything. Until someone runs that test, these are competing advertisements with unusually specific figures.

Primary source, verified: read the paper →

Ground Truth. (2026, September 3). Cerebras is serving an open 27B model at 1,500 tokens a second, and the free tier caps it exactly. Ground Truth. https://groundtruth.day/news/cerebras-serves-an-open-27b-model-at-1500-tokens-a-second.html

@misc{groundtruth:cerebras-serves-an-open-27b-model-at-1500-tokens-a-second,
  title  = {Cerebras is serving an open 27B model at 1,500 tokens a second, and the free tier caps it exactly},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/cerebras-serves-an-open-27b-model-at-1500-tokens-a-second.html}
}

Topics: cerebras · inference · hardware · qwen · open-weight-models · rate-limits