Cerebras

This release for Cerebras Training software introduces support for new models including Llama 3.3 (70B), Llama 3.2 (1B and 3B), and Mistral NeMO (12B). It extends the Max Sequence Length (MSL) up to 128K tokens for training, fine-tuning, and evaluation tasks. Key enhancements include a new Model Zoo Command Line Interface (CLI) that centralizes all modeling tasks, the CSZoo Assistant (a command-line LLM agent leveraging Cerebras Inference), and Pydantic-based Config Classes for streamlined configuration management. Additionally, it offers expanded data preprocessing options (inline, offline, multimodal) and a CS-3+ Performance Upgrade delivering a 1.9x performance improvement over CS-2 systems with linear scaling.

The latest Cerebras Inference Service updates, as of April 2026, bring support for new dedicated models such as GLM 5, GLM 5.1, and Kimi K2.6. A significant performance upgrade introduces speculative decoding, boosting Llama 3.1 70B output speed to an average of 2,100 tokens/second. API requests that fail validation now consistently return HTTP 400 Bad Request instead of 422 Unprocessable Entity. The service now integrates with Microsoft AutoGen, enabling developers to build AI agents with advanced features like tool use and parallel tool calling. Support for OpenAI GPT-OSS (gpt-oss-120b) has been updated with enhanced tool calling (`strict: true`) and expanded JSON Schema limitations, increasing nested levels from 5 to 10 and max properties from 100 to 500. Users are also encouraged to migrate from the `llama3.1-70b` model to `llama-3.3-70b` due to upcoming deprecation.