Zhipu AI's GLM-5.3 now trainable on Fireworks dedicated infrastructure

Zhipu AI's GLM-5.3 model is available for full training through Fireworks' Dedicated Training API and Managed Training surfaces. The model was designed for complex code tasks and long-horizon agents — systems that must plan and execute extended action sequences over long context without losing direction.

The platform supports a range of training methods: supervised fine-tuning (SFT), direct preference optimization (DPO) and odds-ratio preference optimization (ORPO), reinforcement learning (RL) and reward fine-tuning (RFT), distillation from larger to smaller models, and custom loops for cases that fall outside standard patterns. All run on dedicated infrastructure, meaning compute resources are allocated exclusively to your run, not shared with other tenants.

A dedicated training run consumes two separately billed resources: a Trainer that performs optimization steps, and an Inference Deployment used for rollout sampling when evaluation or reinforcement data collection is needed. The basic loop: one or more forward/backward passes, one optimizer step, save the sampler weights, refresh the deployment, then collect new rollouts. Checkpoints are saved at a pre-approved cadence; remaining resources are deleted or scaled down on completion.

Getting started follows ten steps, from installing the SDK and Cookbook through selecting a recipe close to the task, then choosing a model and a Shape — a shared identifier representing a verified training-architecture version linked to a compatible deployment Shape. The SDK automatically resolves the verified version from the full Shape ID passed as training_shape_id. Subsequent steps cover training, checkpointing, promotion, deployment, and cleanup.

The DeploymentSampler interface defines the request contract for RL sampling, with SDK-specific batching behavior; the canonical reference for session affinity is "Inference for RL rollouts." On the loss side, optim_step handles gradient accumulation, with local validation for datum fields, token masks, and normalization. Full loss routing and data details are in the Training API losses skill reference.

The announcement gives no official launch date beyond "now" and provides no benchmarks comparing GLM-5.3 to peer models on code or agent tasks — a gap for developers deciding whether to commit dedicated infrastructure.