Wafer

O que é Wafer?

Wafer provides inference for open models and optimizes how those models are served. Its optimization agent profiles a workload, compares configurations across the model, serving engine, kernels, and hardware, then continues measuring production traffic as conditions change.

For dedicated deployments, teams provide the model, request pattern, and service-level objectives so Wafer can build and tune an endpoint around them. Individual users can access the service through Wafer Pass; its account and usage limits differ from dedicated or enterprise arrangements.

O que Wafer pode fazer?

Workload profiling

Identify bottlenecks in scheduling, decoding, kernels, memory use, and hardware fit using the workload’s actual traffic pattern.

Configuration search

Compare batching, decoding, quantization, engine, kernel, and hardware configurations against the latency or throughput outcome you need.

Kernel and engine optimization

Wafer describes generating fused operations and hardware-specific kernels, and tuning serving-engine settings such as scheduling and the KV cache for the workload.

Dedicated inference endpoints

Build an endpoint around a selected model, real request mix, and service-level objectives, with continued profiling as traffic, models, and hardware change.

Casos de uso

“Latency-sensitive applications”

Evaluate an endpoint against a real application’s traffic and latency requirements, rather than relying only on a generic throughput benchmark.

“Simplifying inference infrastructure”

Test whether faster model responses can reduce application-side workarounds such as prefetching. Wafer presents Brilliant’s removal of its prefetching layer as a customer example.

“Hardware-specific model serving”

Profile and tune a model on the target accelerator stack. Wafer presents its Kimi 2.5 deployment work with DigitalOcean on AMD as an example, rather than a universal model-performance guarantee.

Perguntas frequentes

What does continual inference mean?

Wafer uses this term for profiling workload behavior and repeatedly adjusting the serving configuration. Its described process tests candidate configurations and continues measuring production traffic after deployment.

Can a team share a Wafer Pass account?

No. Wafer Pass is for one individual, prohibits account or API-key sharing and subscription pooling, and permits up to three in-flight inference requests per account. Teams and organizations should arrange a dedicated or enterprise plan.

How does Wafer Pass billing work?

Weekly, monthly, and yearly subscriptions renew automatically unless canceled before renewal. Cancellation stops future renewals while access continues through the paid period. Fees are generally non-refundable, including unused quota; included models, quotas, and overage rates depend on the subscription.

Who owns prompts and model outputs?

Users retain ownership of prompts and completions. Wafer’s terms give it a limited license to process them to operate, secure, and bill the service. Use of outputs must also comply with the served model’s own license; outputs are not guaranteed to be accurate.

Are the customer performance results guarantees?

No. The reported customer outcomes concern particular workloads and deployments. Wafer’s terms do not promise a specific speed, uninterrupted availability, or error-free model output.