Wafer

Wafer is an enterprise LLM platform for serverless and dedicated open-model inference. It helps teams run OpenAI-compatible workloads with low-latency serving, cache-aware billing, and dedicated endpoints for sensitive or performance-critical use cases.

AI Инструменты Разработчика Большие языковые модели

Посетить Сайт

Overview

Wafer is an enterprise LLM platform focused on serving open models through two paths: serverless inference for immediate API access and dedicated endpoints for workloads that need isolation and custom tuning. The homepage describes it as “the fastest open source LLMs for enterprise” and presents it as a way to run open models without managing your own infrastructure.

The public site emphasizes speed, predictable serving, and compatibility with existing AI tooling. Wafer says its Serverless endpoints follow the OpenAI Chat Completions schema, support streaming, tool use, and JSON mode, and can be used with existing clients by swapping the base URL and API key. For more demanding or sensitive deployments, Wafer offers dedicated inference endpoints with workload-specific optimization, SLA-backed uptime, and zero data retention available for compliance-bound use cases.

Core capabilities

Serverless inference

Serve top open models through serverless APIs without managing infrastructure or deployment overhead. Wafer positions this as a pay-as-you-go path for fast access to open models.

Dedicated inference endpoints

Use dedicated endpoints for custom models and mission-critical workloads that need isolation, predictable uptime, and performance tuned to a specific hardware setup.

OpenAI-compatible API

Wafer says its serverless API follows the OpenAI Chat Completions schema, which lets existing OpenAI SDK, LangChain, LiteLLM, Claude Code, and Cline-style clients connect by changing only the base URL and API key.

Agent-friendly response modes

The platform supports streaming, tool use, and JSON mode across Serverless models, which makes it suitable for agent workflows and structured outputs.

Cache-aware billing

Model cards on the homepage expose separate input, output, and cache rates, and Wafer explains that repeated prompt prefixes are billed at the cache rate automatically.

Workload-specific optimization

Wafer publishes benchmark results and states that it profiles models on each accelerator family, shards experts to fit cache hierarchy, and tunes continuous batching for the model’s KV-cache footprint.

Practical use cases

Fast integration for product teams
Teams that want to add open-model inference quickly can use Wafer’s serverless endpoints instead of operating their own serving stack. The site frames this as a pay-as-you-go option with no infrastructure or deployment overhead.
Low-latency conversational systems
Products with voice agents or other real-time interactions can use dedicated endpoints when latency and steady TTFT matter more than raw throughput. The Neon Health case study shows this fit for conversational healthcare workflows.
Compliance-bound enterprise workloads
Organizations handling PHI or other sensitive data can use dedicated endpoints with traffic isolation and zero data retention options. The Neon case study also notes BAA support and US-only data residency for that deployment.
OpenAI-compatible agent stacks
Teams that already use OpenAI-style tooling can connect Wafer with minimal client changes. The site explicitly mentions compatibility with the OpenAI SDK, LangChain, LiteLLM, Claude Code, and Cline-like agents.
Caching-heavy reasoning and coding jobs
Developers running coding, reasoning, or mixed agent workloads can choose among the current Serverless models and use cache-aware billing for repeated prompt prefixes.

Pros and Cons

Pros

Offers both serverless access and dedicated endpoints, giving teams a choice between quick integration and custom deployment.
Uses an OpenAI-compatible Chat Completions schema, which lowers the effort required to switch existing clients.
Supports streaming, tool use, and JSON mode on Serverless models.
Publishes model-level input, output, and cache pricing and explains how automatic cache hits are billed.
Provides a dedicated path for sensitive workloads with traffic isolation, SLA-backed uptime, and zero data retention options.

Cons

The pricing page currently returns a 404, so the public site does not provide a straightforward plan overview there.
The homepage only lists a small set of Serverless models today, with more described as rolling out.
Some of the strongest claims are benchmark-specific, so readers should evaluate them against their own workload shape.

FAQ

Can Wafer work with existing OpenAI-based clients?

Wafer’s Serverless endpoints follow the OpenAI Chat Completions schema, so existing clients can connect by swapping the base URL and API key. The site says streaming, tool use, and JSON mode are supported across Serverless models.

What does Wafer offer for sensitive or regulated workloads?

Wafer says dedicated endpoints isolate traffic from shared inference pools and can be provisioned with zero data retention for compliance-bound workloads. The site also states that Wafer signs DPAs and provides SLA-backed uptime.

Is pricing publicly listed?

The homepage says Wafer’s Serverless offering provides open-model inference with pay-as-you-go access, while dedicated endpoints are for custom model serving and sensitive workloads. The pricing page currently returns a 404, so the public site does not expose a current plan table there.

Which models are available on Wafer?

The site currently lists GLM-5.1, Kimi-K2.6, and Qwen 3.5 397B-A17B on Serverless, and says more models are rolling out. The model cards on the homepage also show cache pricing alongside input and output rates.

What kinds of workloads use dedicated endpoints?

Yes. Wafer’s dedicated endpoint case study for Neon Health says the team moved GLM-5.1 to a dedicated Wafer endpoint and used a TTFT-based SLA to support voice-agent latency requirements.