Tinker
#19 in Reasoning ModelsThinking Machines Lab · since 1. Oktober 2025 (private Beta); 12. Dezember 2025 (General Availability, keine Warteliste mehr) · 20× · last seen Jul 19, 2026
Tinker is a hosted API developed by Thinking Machines Lab (founded by Mira Murati) for fine-tuning open-weight language models using LoRA (Low-Rank Adaptation). It handles distributed training, resource management, and infrastructure reliability, while users retain full control over training algorithms and data via low-level Python primitives (forward_backward, sample). Tinker supports both supervised fine-tuning and reinforcement learning, covering a broad model lineup from compact models to large mixture-of-experts models with over a trillion parameters. Since December 2025 it has been generally available without a waitlist, offering usage-based pricing per million tokens.
Features
| Key-Benchmark (%) | 88,1% pass@32 auf MiniF2F (Theorembeweiser, trainiert via Tinker mit LoRA, nur 20% der Trainingsdaten) |
| Kontextfenster (Token) | Bis zu 262.144 Token (z.B. Qwen3-235B-A22B, Kimi-K2.5:peft); Standard variiert je nach Modell |
| Lizenz | Proprietärer Dienst (Terms of Service); begleitende Tinker Cookbook Bibliothek ist Open Source |
| Multimodalität | Ja, seit Dezember 2025: Vision-Input-Unterstützung mit Qwen3-VL (Bilder, Screenshots, Diagramme neben Text) |
| Plattform | Web-basierte, gehostete Python-API (Cloud-Service) für verteiltes Training auf GPU-Clustern von Thinking Machines |
| Preis pro 1M Token | Ab $0,40 pro Million Token für Training (Beispiel Qwen3-8B); Preise variieren nach Modellgröße und Operation (Training/Sampling) |
| Release-Datum | 1. Oktober 2025 (Private Beta Launch); General Availability ab 12. Dezember 2025 |