Cerebras has put Alibaba’s Qwen 3.8 27B on its wafer-scale inference hardware, and the headline number is absurd in the best way: 1500 tokens per second on a 27B-parameter open model. For context, most cloud providers serve models this size at 30 to 80 tokens per second. Cerebras serves dense open models an order of magnitude faster because its entire chip is the memory-adjacent compute, not a GPU juggling weights between HBM and cores.
It landed on Hacker News the same day, and the discussion is where the real picture emerges.
What you actually get
The practical spec sheet: Qwen 3.8 27B on Cerebras at 1500 tokens per second, a 128K context window, and a public endpoint capped at 150,000 tokens per minute. The 27B class is notable in its own right — this is the successor line to the Qwen 3.6 27B we flagged in July as the local-development sweet spot, and 3.8 continues the pattern: a single dense model that fits on serious consumer hardware but also serves cheaply in the cloud.
At 1500 tokens per second, an entire 128K-context codebase read takes about 85 seconds. A typical coding-agent round trip that takes a GPU cloud 40 seconds of generation takes under 2. That changes what feels usable interactively — it’s the difference between watching the model think and receiving an answer.
The limits are the story
The Hacker News thread was sceptical, and the scepticism is about plumbing, not capability. The 150k tokens-per-minute cap applies to input tokens, which at Cerebras speeds works out to roughly five seconds of actual usage per minute before the meter resets. “Even though each individual request was fast, the sessions we…” ran one developer who tried building on last year’s coding plan and gave up. Another reported hitting the cap on a public endpoint with concurrent agent sessions and concluding the model was “unusable” for real coding workloads at that ceiling.
There’s also a 128K context limit, which commenters flagged as the truer ceiling for coding agents: prompts, files and compaction summaries eat most of it, leaving less room for the actual problem than a long-context cloud model would.
The complaint pattern is familiar from our coverage of cloud AI bill shock — the sticker speed is real, but the rate architecture determines who it serves. Cerebras sells dedicated capacity for anyone who needs more; the public endpoint is a demo with generous limits, not a platform. One commenter’s summary was blunt: “I always want to like Cerebras, but as a tokens-in tokens-out consumer you are not valued at all.”
Why it matters anyway
None of that cancels the underlying shift. A 27B open model served at 1500 tokens per second means the capability-per-dollar frontier for interactive inference has moved decisively. We’ve tracked Qwen’s run all year — 3.6 Plus passing a trillion tokens on OpenRouter, 3.6 27B beating much larger closed models on coding, and the 3.8 Flash preview pointing at the next generation. The pattern is consistent: Alibaba ships open weights, and hardware vendors race to serve them as fast as physics allows.
For anyone running AI infrastructure in New Zealand — an agency, a startup, a university lab — the calculation is straightforward. Cerebras access isn’t local, and rate limits make the public tier unsuitable for sustained agent workloads. But the direction matters: open models one generation behind the frontier, served at speeds that change product design. If you’re building anything where response latency is the bottleneck, watch this space rather than the benchmark leaderboards.
FAQ
How fast is Qwen 3.8 27B on Cerebras? About 1500 tokens per second on the public endpoint — roughly 20 to 50 times faster than typical GPU cloud serving of a 27B model.
What are the limits? A 150,000 tokens-per-minute cap on the public endpoint (counting input tokens) and a 128K context window. Cerebras offers dedicated capacity above those limits for enterprise customers.
Can I run Qwen 3.8 27B locally? Yes — it’s an open-weight 27B dense model, which puts it in quantised range for high-end consumer GPUs and comfortably within reach of workstation-class Apple Silicon. Cerebras is the fast path; local hardware is the private path.
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.
