Fireworks AI

Fireworks is an AI platform that provides training and inference tools to transform open models into domain-specific or specialized intelligence. The service offers capabilities for model training, fine-tuning and serving to support customized AI deployments.

Fireworks AI's training API is a programmatic model-training offering for ML engineers who need to define custom training loops and loss functions beyond managed training. The video describes support for supervised fine-tuning, reinforcement fine-tuning, and direct preference optimization (DPO), as part of Fireworks AI's train-and-serve workflow for open models.

Fireworks AI's inference engine for serving AI models. It uses custom-built CUDA kernels that can be optimized for a customer's workload to target latency and performance requirements. The video credits FireAttention and related Fireworks serving work with enabling 13× faster inference for Cursor's Composer 2.

A managed fine-tuning service from Fireworks AI that lets users select a supported base model, provide training data and configuration, and launch jobs through the Fireworks UI, CLI, or API without writing custom training infrastructure. Fireworks-managed infrastructure handles scheduling, training, checkpointing, and model output; the same jobs can also be configured with an agent. Training data uses the OpenAI-compatible chat-completion format, and the UI provides visual monitoring of training metrics. Managed training supports supervised fine-tuning for text and vision models, preference optimization with DPO or ORPO, and LoRA adapter training. Supported model families include DeepSeek, Qwen, Kimi, Gemma, GLM, and Llama, with eligibility and context limits determined per model and method. Per-step loss and learning-rate artifacts are retained for seven days unless longer-term history is configured through Weights & Biases. Full-parameter tuning is available through Fireworks' separate Training API; managed reinforcement fine-tuning is deprecated.

Fireworks Training Agent is a UI-accessible fine-tuning tool from Fireworks AI. It accepts a dataset and a description of the training goal, then selects a base model, cleans the data, plans and runs the fine-tuning job, chooses hyperparameters, and reports the results. Fireworks describes this guided training path as letting users describe the task, review the plan and cost, approve the run, and receive a trained model; the demonstrated workflow used supervised fine-tuning on Qwen 3 8B.