Cerebras

Cerebras Systems, founded in 2015 and based in Sunnyvale, California, builds wafer-scale AI hardware and the cloud and software platform around it. Developers can call hosted open and frontier models through an OpenAI-compatible API, while organizations can also train or fine-tune models or run systems on-premise. Cerebras says its inference is up to 30x faster than GPU systems, and notes that comparisons rest on third-party benchmarking or internal testing and vary by workload.

What It Is

Cerebras is an AI inference and training platform built on its Wafer-Scale Engine (WSE) chips and CS-series systems. Access comes through the Cerebras Inference Cloud, which offers a public endpoint with an API key. The site also lists dedicated and on-premise deployment options. The homepage lists models such as Qwen, GPT 5.6 Sol Ultrafast, Kimi K2.6 and Codex-Spark. It states that inference runs in-region to help with latency, data residency and compliance requirements.

Software Stack and Training

The Cerebras Software Platform (CSoft) is aimed at training large models without writing distributed-computing code. Its documented components are:

• PyTorch support through a lightweight API wrapper that uses XLA's lazy tensor backend.
• The Cerebras SDK and Cerebras Software Language (CSL), for writing custom kernels with a C-like interface.
• The Cerebras Graph Compiler, which turns a neural network into an executable program for the WSE-3 by allocating cores and minimizing communication latency.
• AI Model Studio, a pay-per-model training service on dedicated CS-3 clusters hosted by Cirrascale Cloud.

Hardware and Update

The company page says the CS-4 launched in August 2026. It is powered by three WSE-3 Turbo processors and is described as delivering up to 30x faster inference than GPU systems. The same page says Cerebras began trading on Nasdaq under the ticker CBRS in May 2026. The homepage also announces a disaggregated inference partnership with AMD and OpenAI's GPT-5.6 Sol running in Ultrafast mode on Cerebras at up to 750 tokens per second.

Typical Use Cases

The site highlights low-latency coding assistance, multi-step agent workflows, fast reasoning for search and copilots, and voice conversations. Customer case studies on the site include Cognition, NinjaTech, AlphaSense and Tavus.