Qwen 3.8 27B Model Now Available on Cerebras at 1500 Tokens/s
Cerebras offers Qwen 3.8 27B with a processing speed of 1500 tokens per second.
1Features of Qwen 3.8 27B
Qwen 3.8 27B is an AI model with 27 billion parameters. It's accessible on the Cerebras platform through free trial and pay-as-you-go options. The model offers 64k free and 128k paid context windows. It stands out with a processing speed of 1500 tokens per second. This performance demonstrates it's faster than many models of similar size.
2Capabilities of Cerebras Platform
Cerebras provides access to various model families for its users. The platform also hosts models like OpenAI GPT OSS with 120 billion parameters. Users can select suitable models based on their needs using the model selection guide. Dedicated endpoints offer higher throughput and production SLAs. The platform facilitates new users with a quickstart guide.
3Model Compression Technologies
Cerebras hosts original, unpruned versions of all models on its platform. Weights are stored in partial 16-bit, 8-bit, and 4-bit formats following industry standards. Sensitive layers are stored at full precision while operations are performed at high precision. Although researching techniques like REAP, these pruned models are currently only shared with the research community on Hugging Face.
4Processing of Quality Layers
Activations, attention, and kv cache remain at full precision without quantization on the platform. This approach maintains maximum model quality. Weight-only quantization is applied only during storage. High-precision calculations are performed during operations. This method reduces storage costs without negatively affecting model performance.
5Use Cases and Limitations
Users can select model names to view complete specifications, capabilities, and per-tier limits. The platform offers a model selection guide for different use cases. The free tier has certain rate limitations. Dedicated endpoints are recommended for higher throughput and production SLAs. Users can quickly get started by making API calls.
6REAP Pruning Technique
Cerebras conducts research on pruning techniques like REAP (Router-weighted Expert Activation Pruning). This technique aims to reduce model size while maintaining performance. However, pruned models are currently only shared for research purposes. These models aren't available through the public API. Detailed information about REAP can be found in the research blog.
7Frequently Asked Questions
The platform addresses users' frequently asked questions. It clarifies whether model architectures might change without notice. Technical details like where to find REAP pruned models are shared. Users can thoroughly examine model specifications and limitations on the platform. API documentation simplifies integration processes.
8Contribution to AI Ecosystem
Cerebras making models like Qwen 3.8 27B accessible significantly contributes to the AI ecosystem. Easy access to high-performance models offers great opportunities for researchers and developers. The platform's transparent approach helps users fully understand model characteristics. These developments help accelerate innovation in AI. Cerebras' solutions are raising industry standards.