What differs
- Control: Tinker exposes low-level training primitives so your code defines the loop; CoreWeave offers Serverless RL with your reward function and Serverless SFT (public preview), recommended through the ART framework.
- Pricing: Tinker publishes per-token prefill, sample, and train rates per model (for example GPT-OSS-120B train $0.737 per 1M tokens at 32K context); CoreWeave bills training per GPU-hour of active time, with the rate on its pricing page.
- Weights: Tinker offers an API endpoint to download saved checkpoints; CoreWeave stores LoRA adapters as Weights & Biases Artifacts that can be saved locally.
- Serving: Tinker lists serverless inference in beta; CoreWeave hosts every trained checkpoint on Serverless Inference.
- Ownership and compliance: CoreWeave is publicly listed and acquired OpenPipe in 2025; it lists SOC 2, ISO 27001, 27017, and 27018 in its compliance programs. No certification was found on Tinker’s site.
Who each tends to fit
Researchers who want to change the training algorithm itself may look at Tinker. Teams training agents with RL that already use Weights & Biases may look at CoreWeave. Both use LoRA only.
Last checked: . No winner declared.