Test your knowledge in live multiplayer Jeopardy-style trivia battles with AI-assisted grading that recognizes correct answers even with spelling mistakes.
castform
Train and deploy state-of-the-art open models with serverless inference, synthetic generation, and usage-based pricing on token rollouts.
Free plan available
$0.03/1m tokens
About castform
castform is an AI platform designed for training open-weights models and running serverless model inference. The platform enables machine learning teams and software developers to execute fine-tuning and post-training runs on supported base architectures, metering expenses entirely on the tokens processed during rollout generation and model weight updates. In addition to training facilities, castform hosts serverless endpoints that support both base model queries and direct serving of trained checkpoints. The service provides a flexible infrastructure supporting modern model architectures such as Qwen3-VL, Qwen3.5, Gemma 4, and Nemotron 3.5, offering up to 256k context windows. Its system architecture includes a synthetic data generation pipeline and judge rewards mechanisms, which integrate with high-end models like GPT-5.6, Grok, and Step architectures to facilitate complex reinforcement learning and evaluation workflows. Checkpoints can be immediately served via API endpoints without complex server setup. castform is primarily built for machine learning engineers, AI researchers, and technical development teams looking to fine-tune open models without managing on-premise or cloud GPU clusters. It is also suitable for organizations needing a single unified provider for synthetic data pipelines, evaluation rewards, and production deployment of customized models. The platform offers transparent token-based pricing alongside an initial credit balance to lower the barrier to entry. What sets castform apart is its cohesive workflow combining reinforcement learning rollouts, synthetic data generation pipelines, and direct checkpoint serving. Instead of juggling separate orchestration, GPU rental, and inference vendors, users can initiate runs on foundation models, apply judge rewards, and transition models directly into low-latency serverless inference under consistent, metered token rates.
castform pros & cons
Pros
- Provides $50 in free credits to test training runs without upfront payment.
- Charges the same inference rates for custom fine-tuned checkpoints as base models.
- Supports long context lengths up to 256k across multiple open-weight models.
Cons
- Very large models like Step 3.7 Flash and Hy3 require contacting sales rather than self-serve access.
- Certain URLs such as the xcode documentation route return 404 errors.
castform use cases
- AI researchers can train vision-language models like Qwen3-VL using up to 256k context windows for document analysis.
- Machine learning engineers can generate synthetic datasets and use judge reward models to fine-tune open models with reinforcement learning.
- Backend developers can immediately deploy and serve fine-tuned model checkpoints via serverless inference without managing GPU instances.
castform features
- initial $50 free credit grant for new users
- embeddings generation endpoints via models like text embedding 3 large
- support for models including qwen3, gemma 4, and nemotron 3.5
- judge rewards integration for reinforcement training
- integrated synthetic data generation pipeline
- support for large context windows up to 256k
- serverless model inference for base models and custom checkpoints
- usage-based model training metered per 1m rollout and update tokens
castform pricing
Is castform free? Yes, castform has a free plan, and paid plans start at $0.03 / 1m tokens.
Free Credits
Free
- $50 in free credits to start your first training run
- Access to open-source base models for training
- Serverless inference testing
Usage-Based Inference
$0.03 / 1m tokens
- Identical pricing for base models and trained checkpoints
- Input pricing starting at $0.03 per 1M tokens
- Output pricing starting at $0.15 per 1M tokens
- Text embedding models from $0.143 per 1M input tokens
- Serverless serving of trained model checkpoints
Usage-Based Training
$0.10 / 1m tokens
- Metered on rollout generation and weight update tokens
- Rates starting at $0.10 per 1M training tokens for small models
- Rates starting at $0.20 per 1M training tokens for medium models
- Support for up to 256k context length
- Synthetic generation pipeline and judge rewards
castform FAQs
How is training metered on castform?
Training is billed in USD per 1M tokens based on what your training run processes. This includes the tokens used for rollout generation as well as weight updates.
Are there free credits available to try castform?
Yes, castform provides $50 in free credits to help you get started with your first training run.
Does inference cost more when serving trained checkpoints?
No, inference pricing on castform is identical whether you are querying a base foundation model or serving your custom trained checkpoints.
What context length is supported for model training?
Training models supported on the platform, including Qwen3-VL, Qwen3.5, and Gemma 4, support up to a 256k context window.
Ratings & reviews
No reviews yet. Be the first to share how castform worked for you.
Sponsored tools
Connect with engaged niche audiences or monetize your subscriber base through an automated marketplace featuring verified metrics and secure Stripe payments.
Swap characters, backgrounds, or styles in any video while preserving original motion and camera paths. Perfect for digital marketers and creators.
Track brand citations and mentions across Google AI Mode, ChatGPT, Claude, and more. Discover top-cited sources and queries with zero signup required.