Fireworks AI released a training platform in preview that supports full-parameter fine-tuning for models ranging from 8B to 1T parameters. This allows teams to move beyond prompt engineering by using reinforcement learning to build proprietary models that outperform closed frontier systems on specific tasks.
Fireworks AI, an inference platform for fast model serving (generating outputs from trained models), launched Fireworks Training in preview. The platform supports full-parameter fine-tuning for models up to the 1-trillion-parameter Kimi K2.5. It offers three interfaces: a conversational Training Agent, managed infrastructure, and a Training API.
This shift addresses the performance ceiling teams hit with prompt engineering on closed models. By fine-tuning open-weight models like Qwen3 using reinforcement learning, companies are achieving higher accuracy and lower latency than proprietary systems. The platform also ensures numerical parity so models behave identically during training and production inference.
You can now use the Training API for custom algorithms like GRPO or the Training Agent to automate hyperparameter sweeps. Multi-LoRA serving is also available, allowing you to run multiple fine-tuned adapters on a shared base model. The platform is in preview for teams building proprietary moats.
