Tinker
#9thinking-machines-lab · seit 2025-10-01 · 16× · zuletzt 30. Juni 2026
44
Momentum
Tinker is a managed Python API by Thinking Machines Lab (founded by former OpenAI CTO Mira Murati) for fine-tuning open-source language models via LoRA. The API abstracts all infrastructure—scheduling, GPU resource allocation, and failure recovery—allowing researchers and developers to write custom training loops locally in Python while Thinking Machines provides the distributed GPU compute. Supported models range from 1B to 1T+ parameters (dense and MoE), including Qwen3, Kimi-K2, and GPT-OSS variants. The product launched on October 1, 2025, and has been generally available without a waitlist since December 2025.
Historique du momentum
04.04.03.07.
Fonctionnalités
| License | Tinker is a commercial managed service (proprietary). The companion library Tinker Cookbook is open source (publicly available on GitHub under thinking-machines-lab/tinker-cookbook). |
| Price per 1M Tokens | $0.40/M tokens for Qwen3-8B training (example price; full pricing model-dependent on the official page) |
| Release Date | October 1, 2025 (Private Beta); General Availability from December 12, 2025 (waitlist no longer required) |
