a training run is where your environment and dataset come together: castform runs the rl training loop on gpus it provisions for you, and you end up with a fine-tuned model to evaluate and deploy. this page covers what happens during a run and which base models you can train.
what happens during a training run
- your environment class and datasets are uploaded to castform’s storage
- GPUs are provisioned (takes a few minutes)
- the training loop begins: the model generates completions, your reward function scores them, and the model updates its weights
- metrics and model outputs are reported to the console as the run progresses (updated periodically, not streamed live)
supported models
the platform currently supports four base models for fine-tuning. pass the model id via launcher_args.model when launching a run. per-token training rates for each model are on the pricing page.
| model | description |
|---|---|
Qwen/Qwen3.5-4B | lightweight 4B parameter model. fast iteration, lower GPU cost. good default for experimentation. |
Qwen/Qwen3-VL-4B-Instruct | 4B vision-language model. the choice for multimodal environments with image prompts and tools. |
Qwen/Qwen3.5-35B-A3B | 35B mixture-of-experts model with 3B active parameters. stronger baseline, similar inference cost to the 4B. |
google/gemma-4-26B-A4B-it | 26B mixture-of-experts model with 4B active parameters, from google. the option when you need a US-developed model. |