Fireworks AI Adds Reinforcement Learning for GLM 5.1 to Build Custom Agents

May 15, 2026

Fireworks AI, an inference platform for fast model serving, added reinforcement learning (RL) support for GLM 5.1 via its Training API. This update builds on the platform's GLM 5.1 fine-tuning support by using LoRA (an efficient adaptation technique) to perform RL on a 200K context window.

Context window: 200K tokens
Training methods: SFT, DPO, and LoRA RL
Usage limits: None
Model ownership: Full weight ownership
Access: Training API

Adding RL matters for teams building specialized reasoning agents that fine-tuning alone can't shape. Custom loss functions let teams encode their own domain objectives instead of relying on rigid recipes, turning proprietary data into a differentiated model. This follows the launch of Fireworks AI's Qwen 3.6 fine-tuning and a broader expansion of the Fireworks Training Platform.

You can access these capabilities through the Training API with no usage ceilings. SFT, DPO, and RL run on the same infrastructure that serves Fireworks production inference, so a trained checkpoint becomes a live endpoint without format conversion or stack migration. You retain ownership of the resulting model weights for your own deployment.

View the full update on fireworks.ai

Fireworks AI

@FireworksAI_HQ4d ago

Fireworks Training Platform continues to expand. Today GLM 5.1 LoRA RL is now live via Training API: SFT, DPO, and full RL on a 200K context window → custom loss functions or smart defaults. No usage ceilings. No credits to claim. Your model. Your inference. Get started → https://t.co/sBNnhKT5dq

220

View on X

Still wondering? A few quick answers below.

GLM 5.1 is a large language model developed by Z.ai. It is available on the Fireworks AI platform for customization, allowing developers to use reinforcement learning to adapt the model to their own data and reward signals. The Fireworks Training Platform handles the infrastructure so teams can focus on the training objective rather than GPU orchestration.

The Fireworks Training Platform supports reinforcement learning for GLM 5.1 using Low-Rank Adaptation, or LoRA, which is an efficient method for fine-tuning large models. Developers can use the Training API to implement supervised fine-tuning, direct preference optimization, or full reinforcement learning. The platform allows for custom loss functions or the use of smart defaults during the training process.

Fireworks AI provides access to GLM 5.1 training with no usage ceilings or credit-claiming requirements. This means developers can train and deploy their personalized models without hitting the restrictive caps often found on closed-source frontier platforms. Once training is complete, the custom model weights are available for inference on the same Fireworks infrastructure.

Training for GLM 5.1 on the Fireworks platform supports a 200K token context window. This large window allows the model to process and learn from extensive datasets, such as long technical documents or codebases, during the reinforcement learning and fine-tuning phases. This capability is useful for tasks that need coherence across substantial amounts of context.

Yes, the Fireworks Training Platform is designed so that you own your model and your inference. After using the Training API to customize GLM 5.1 with reinforcement learning or fine-tuning, the resulting model is yours to use. You can run inference on your personalized model directly through the Fireworks cloud, which is optimized for fast and reliable generative AI performance.