Qwen 3.8 27B runs at 1500 tokens per second on Cerebras inference endpoints

Qwen 3.8 27B at 1500 tokens per second

A front-page Hacker News post reports that Qwen 3.8 27B is now available on Cerebras' inference service at roughly 1500 tokens per second. The post links to Cerebras' model catalog, the documentation page that lists everything served on the company's public endpoints and explains how those models are hosted.

That figure is the story. A 27-billion-parameter open-weight model generating on the order of 1500 tokens per second through a shared API is an unusually high generation rate, and Cerebras, which builds wafer-scale AI accelerators rather than conventional GPUs, has made raw inference speed its central pitch. Adding a recent open-weight Qwen release extends that pitch to developers who want frontier-adjacent open models rather than only the small or mid-size options that speed-focused platforms typically showcase.

What the model catalog covers

According to Cerebras' documentation, models on the public endpoints can be used on both the free trial tier and pay-as-you-go, with rate limits and pricing applying. Reserved capacity, higher throughput, production service-level agreements and additional model families are offered through dedicated endpoints instead.

The page also spends considerable space on compression, an unusual level of detail for a model listing and one clearly aimed at developers who want to know exactly what they are calling.

Original architectures, storage-only quantization

Cerebras says it does not host pruned models on its public endpoints; every model there is described as the original, unpruned version. The company does research pruning under the name REAP (Router-weighted Expert Activation Pruning), but those pruned variants are shared with researchers on Hugging Face rather than served through the shared API.

Quantization is treated separately. Weights are stored with selective weight-only quantization in a partial mix of 16-bit, 8-bit and 4-bit precision, while layers considered sensitive stay at full precision and are dequantized as they are loaded so that operations execute in high precision. Activations, attention computation and the KV cache are kept in full precision and left unquantized, per the documentation.

The company also commits to leaving served architectures untouched: it will not modify a model's architecture without notice, and any future pruned offerings would appear as separate endpoints carrying pruning-specific names so developers can choose which version to use.

Why it matters

Speed changes what developers can build. Agent loops, multi-step code generation and long-form drafting are all constrained by how quickly tokens come back, and a 1500-tokens-per-second rate turns those workloads from slow batch jobs into interactive ones, without anyone having to buy and operate hardware.

The transparency commitments matter too. A common concern with hosted open-weight models is that the served version quietly differs from the published weights. By pledging unpruned architectures and confining quantization to storage, Cerebras gives developers a clearer contract about what the API actually runs.

Finally, fast hosted inference for an open-weight model keeps competitive pressure on closed-model APIs: developers get speed without giving up the ability to benchmark, inspect or, if necessary, host the weights themselves.