Cerebras Inference Offers GPT OSS and Qwen Models on Public Endpoints

Cerebras has expanded its Inference public endpoints, providing developers with direct access to a streamlined catalog of high-performance generative models under flexible deployment tiers. The platform currently hosts OpenAI GPT OSS, identified as gpt-oss-120b, alongside Qwen 3.8, listed as qwen-3.8-27b. Accessible through free trial and pay-as-you-go options subject to standard rate limits, these models can also be deployed via Dedicated Endpoints for organizations requiring reserved capacity, higher throughput, and production service level agreements. Performance benchmarks for the current catalog indicate strong inference speeds. The 120-billion-parameter GPT OSS model processes approximately 3,000 tokens per second, supporting context windows of 65,000 to 131,000 tokens depending on the selected tier. The 27-billion-parameter Qwen 3.8 model delivers roughly 1,500 tokens per second with context limits ranging from 64,000 to 128,000 tokens. These specifications position the platform as a competitive solution for latency-demanding enterprise applications. Cerebras has implemented a strict compression and architecture policy to maintain inference fidelity across its production network. The company explicitly confirms that all public endpoints serve original, unpruned model weights. Storage optimization utilizes selective weight-only quantization, with parameters stored in partial 16-bit, 8-bit, or 4-bit formats. To preserve accuracy, computationally sensitive layers remain in full precision and undergo dynamic dequantization during execution. Additionally, activations, attention layers, and key-value caches are maintained in unquantized state throughout the inference process. Research into advanced compression remains separate from production services. Cerebras is actively developing Router-weighted Expert Activation Pruning, known as REAP, to reduce computational overhead without degrading model quality. While REAP variants have been published on Hugging Face for academic experimentation, they are excluded from the production API. The company has committed to complete transparency regarding future compression deployments, promising that any pruned architectures will be offered as distinctly named endpoints to allow users to choose between baseline and optimized versions. By isolating experimental compression techniques from its core inference infrastructure, Cerebras ensures predictable, high-precision performance for commercial workloads. The updated endpoint catalog and storage policies provide developers with a stable, scalable environment for integrating large language models into production pipelines.