Tinker

Tinker is a training API for researchers built by Thinking Machines Lab (Mira Murati's laboratory). It exposes a small, composable API surface for fine-tuning and training foundation models — four functions (forward_backward, optim_step, sample, save_state) give full control over the training algorithm, while TML handles infrastructure, scheduling, and GPU-cluster resilience. Aimed at AI researchers and engineers who want control over data and algorithms but not the cluster.

Supported models (as of July 2026): Inkling (975B/41B active, hybrid + vision + audio), Inkling-Small (soon), DeepSeek-V3.1, Kimi K2.6 (K2.5 retiring), NVIDIA Nemotron 3 (Nano 30B / Super 120B / Ultra 550B), OpenAI GPT-OSS (20B/120B), Qwen3.6 (35B-A3B, 27B), Qwen3.5 (4B/9B/35B-A3B-Base/397B-A17B) and Qwen3-8B. Each model is available in several context-window sizes (32K/64K/128K/256K). MoE priced by active parameters — significantly cheaper than dense models of comparable quality.

Training techniques: SFT (supervised fine-tuning), RL (Reinforcement Learning) with loss functions PPO/CISPO/DPO/DRO/cross-entropy/importance-sampling/custom, preference learning (DPO, RLHF), distillation (SDFT, on-policy and off-policy). Tinker uses LoRA as the default fine-tuning technique — an adapter is trained instead of full weights, giving quality comparable to full fine-tuning at significantly lower compute cost (TML research on the blog). Every saved checkpoint can be downloaded via API and moved to your own inference infrastructure (HuggingFace, SGLang, vLLM, llama.cpp).

Ecosystem: open-source Tinker Cookbook (github.com/thinking-machines-lab/tinker-cookbook) contains real recipes — Chat SL, Math RL, Code RL, Preference, Search-R1 Tool Use, Prompt Distillation, Multi-Agent RL, VLM Classifier, RLHF Pipeline, True-Thinking Score and many more. Sampling APIs compatible with both the OpenAI and Anthropic APIs — you can drop a trained model into an existing client without changing code. The Inkling Playground in the Tinker console provides a chat interface with integrated web search. Academic users: Berkeley (Sky Lab), Princeton (Goedel Prover), Stanford (StatMech), Redwood Research.

Pricing (effective from 17 July 2026): billing per million tokens, separate rates for prefill, sample and train. 80% discount on cached prefill. Example train rates (per 1M tokens, prices effective 17 Jul): Inkling 64K $5.61, Inkling 256K $11.23, Nemotron 3 Ultra 64K $5.48, Nemotron 3 Super 64K $1.28, Qwen3.5-4B $0.74, GPT-OSS-20B $0.40. Storage $0.10 per GB-month. No free tier — pay-per-use. User training data is used solely to fine-tune the user's model, not to train TML's own models.