Cerebras

The chip generation behind the service is the WSE-3, and Cerebras runs inference on the Wafer-Scale Engine instead of on many small dies linked by a network. Keeping compute and memory on one piece of silicon removes the cross-chip hops that slow clustered GPUs. The company reports inference that is up to 30 times faster than GPU systems, and results vary by workload and model.

The endpoint at api.cerebras.ai/v1 mirrors the OpenAI client, so existing code usually works once you change the base URL and the model name. Python and Node.js SDKs ship from Cerebras, and cURL calls work with no install at all. This is the feature that makes switching cheap.

Shared Inference runs the public catalog on demand, with per-model rate limits and no setup. Dedicated Inference reserves capacity for one organization and adds stable endpoints, configurable throughput, custom weights, a Management API, a Batch API, and Prometheus-compatible metrics.

Models expose a reasoning_effort setting with none, low, medium, and high levels, so you can dial thinking up for a hard task or turn it off for a simple one. A reasoning_format option decides whether thinking tokens are returned separately, prepended to the answer, or hidden.

Limits are measured in requests and tokens per minute, hour, and day, whichever comes first. Cached tokens skip the uncached limit, so a high cache hit rate can raise effective throughput several times over within the same quota. Developer accounts drop the hourly and daily caps entirely.

Enterprise users define endpoints, deployments, and replicas through a Management API. You deploy fine-tuned weights beside standard versions, tune capacity and quantization, and route requests by model architecture, all without changing the endpoint ID that your app already calls.

Beyond the two shared models, Dedicated Inference supports Qwen3 and Qwen3-Coder, GPT OSS, Llama 3 and 4, Mistral and Devstral, GLM 4 and 5, Kimi K2, DeepSeek V3, Gemma 4, and others. Multimodal models such as Gemma 4 accept image input, and each model can be tuned for latency or throughput.

The Batch API runs large asynchronous workloads on reserved capacity, which suits jobs that do not need an instant answer. It pairs with predicted outputs, a setting where you supply the expected content to cut latency further.