Contents
In brief
Cerebras lists Qwen 3.8 27B (qwen-3.8-27b) on its public inference endpoints at about 1500 tokens/s.
Context is 64k on the free tier and 128k on paid. On the shared surface the company promises original, unpruned weights.
REAP research stays on Hugging Face, outside the production shared API. For teams comparing cloud inference, that is two signals at once.
First is latency and interactivity. Second is which architecture sits behind the model id—not only a marketing family name.
What happened
The Cerebras models overview now includes Qwen 3.8 27B alongside GPT OSS 120B (~3000 tokens/s).
The model has 27 billion parameters and a stated speed near 1500 tokens per second. Context follows the same free/paid pattern with different limits.
Public endpoint models are available on free trial and pay-as-you-go, subject to rate limits and pricing.
For reserved capacity and production SLAs, the docs point to Dedicated Endpoints. The public shelf and the “production contour” are explicitly separated.
A dedicated compression section answers what many vendors leave fuzzy: what actually runs behind the endpoint.
Cerebras says it does not host pruned models on public endpoints. Everything on the shared surface is the original architecture.
Research like REAP (Router-weighted Expert Activation Pruning) is published on Hugging Face, but not served through the production shared API.
That matters for regressions: if you locked quality on qwen-3.8-27b, you should not silently slide onto a thinned checkpoint under the same marketing name.
For storage, Cerebras uses selective weight-only quantization: partial 16-/8-/4-bit weights, with sensitive layers kept at full precision and dequantized on the fly.
Activations, attention, and the KV cache stay unquantized. The FAQ commits to not changing architecture without notice.
If pruned variants appear, they will ship as separate endpoints with explicit names—so clients choose the size/quality trade-off themselves.
Why it matters
For teams running LLMs in chat, agents, and code-generation pipelines, four-digit token rates change product feel.
Less waiting on long answers, smoother interactive loops, cheaper-feeling batch jobs. In practice that also affects user wait cost.
A separate effect is how many parallel sessions one API key can sustain under the same limits. Speed without compression transparency is still a trap.
Pruning alters architecture. Silent weight swaps break benchmark reproducibility and regression suites.
Cerebras’s split is pragmatic: public API = unpruned original; experimental slices = Hugging Face.
Weight-only quantization for storage is framed separately from pruning: save memory without rewriting topology on existing endpoints.
For engineers, the takeaway is to read compression policy, tier context limits, and the exact Model ID—not only tokens/s.
Otherwise it is easy to confuse a Hugging Face card with what actually answers in the cloud.
Also watch free-tier quotas: “fast on a demo” and “stable under load” are different modes of the same public shelf.
| Check | Public endpoints | Research slices |
|---|---|---|
| Architecture | Original, unpruned | May include REAP and others |
| Where | Cerebras docs / API | Hugging Face |
| Context | 64k free / 128k paid | Depends on model card |
| Speed (stated) | ~1500 tokens/s for Qwen 3.8 27B | Not the production API |
In practice
Lock a quality contract before you migrate production traffic for a cheaper token price.
Otherwise a speed win turns into surprises on long context, strict JSON schemas, or tool chains.
Decide early what matters most: max speed, closest match to an open checkpoint, or predictable service level.
If you compare against a local run, separately test long context and tool-using flows—those are where “model card” and cloud serving diverge most often.
- Confirm
qwen-3.8-27bin the docs: context limits (64k / 128k) and real tier quotas. - Run your own regressions (instructions, JSON schemas, long context)—do not trust third-party speed tables alone.
- Do not confuse Hugging Face REAP (and other research) weights with what Cerebras’s production API serves.
- Remember storage quantization ≠ pruning; for a byte-for-byte local checkpoint match, clarify the served format.
- For SLAs and steady throughput, look at Dedicated Endpoints.
After the pilot, record the exact Model ID, tier, and verification date in the ticket—otherwise nobody remembers which contour you used for regressions a month later.
The public shelf is fine for a pilot. Production with hard latency and quota SLAs usually needs the reserved contour.
Takeaway
Qwen 3.8 27B on Cerebras is fast public access to a 27B model with an unusually clear compression story.
Originals on the API, pruning in research, weight quantization as a memory trade-off.
For production, both ~1500 tokens/s and knowing which model sits behind the endpoint matter—otherwise paper benchmarks and live behavior drift apart.