Tinker
Training API for LoRA fine-tuning and sampling of open-weight models, including Thinking Machines' own Inkling family, with the training infrastructure run by Thinking Machines.
What it does
- Model training
- Model inference
Features 17
- forward_backwardAPI call that runs a forward and backward pass and accumulates the gradient.
- optim_stepAPI call that updates trainable weights based on the accumulated gradient.
- sampleAPI call that generates tokens for chat, evaluation or RL rollouts.
- save_stateAPI call that saves weights and optimizer state so training can resume later.
- Managed GPU clustersScales from a few examples to a full training set without coordinating GPUs or handling hardware failures.
- Open-weight model supportSupports open-weight dense and MoE models from 4B to 1T parameters.
- Training recipesStarter recipes include Math RL, Terminal agents, Async RL, supervised fine-tuning, RLHF and on-policy distillation.
- LoRA fine-tuningImplements low-rank adaptation (LoRA) fine-tuning rather than full fine-tuning.
- Vision-language model supportSupports vision-language models such as Qwen3.6-35B-A3B for image understanding tasks.
- Checkpoint weight downloadLets you download the weights of a trained model to use outside Tinker.
- Session metricsShows live utilization, throughput and in-flight requests for every running session.
- Perfetto tracesProvides request-level traces with latency percentiles and rate-limit events.
- Audit logRecords checkpoint, project, team and organization events for admins.
- Customer-managed encryptionLets customers protect their data with their own Cloud KMS solution, managed entirely by them.
- Periodic checkpointsSaves checkpoints automatically at configured training-step intervals.
- OpenAI-compatible inferenceLets you interact with any model checkpoint in Tinker through an OpenAI-compatible API.
- Anthropic-compatible inferenceLets you use any model in Tinker through an Anthropic-compatible API.
How it compares
-
Gemini Enterprise Agent Platform (formerly Vertex AI)
AWS Bedrock, Microsoft Azure AI, and Google Vertex AI provide model-agnostic platforms that let enterprises access multiple foundation models through unified APIs. These platforms compete directly with Thinking Machines' customization layer, offering similar fine-tuning and deployment capabilities across multiple model providers rather than being locked into a single foundation model.
Research report · 22 Oct 2026 -
Amazon Bedrock
AWS Bedrock, Microsoft Azure AI, and Google Vertex AI provide model-agnostic platforms that let enterprises access multiple foundation models through unified APIs. These platforms compete directly with Thinking Machines' customization layer.
Research report · 19 Sep 2026 -
Microsoft Foundry
AWS Bedrock, Microsoft Azure AI, and Google Vertex AI provide model-agnostic platforms that let enterprises access multiple foundation models through unified APIs. These platforms compete directly with Thinking Machines' customization layer.
Research report · 19 Sep 2026 -
Hugging Face Jobs
Hugging Face has built the dominant open-source ecosystem for model deployment and fine-tuning, while NVIDIA's AI foundry services target enterprises wanting custom models.
Research report · 19 Sep 2026 -
NVIDIA AI Enterprise
NVIDIA's AI foundry services target enterprises wanting custom models.
Research report · 19 Sep 2026
Sources
- Pricing
-
Pricing page: "Training & sampling prices in US dollars per million tokens, as of October 19, 2026"; storage is charged per GB per month; serverless inference for Inkling is labelled Beta (not quoted).
- Description
- Constraints
-
Official documentation · 18 Oct 2026
Derived from the security facts on this page, from the same source.
- Security
-
Official documentation · 18 Oct 2026
Thinking Machines Lab Terms of Service: 'we will not use your Customer Content for model training or fine-tuning' and will not share it with a third party to develop their technologies. No trust centre or certification is published (the Security link is a vulnerability-disclosure policy), so certifications stay empty; retention and residency not stated.
- Sold within
-
Tinker's "Models & Pricing" page states "All prices are per million tokens. Checkpoint storage is charged at $0.10 per GB per month." and "We provide an 80% discount on cached prefill tokens.", with Prefill, Sample and Train rates per model. The product page offers a self-serve "Sign up" and says "Tinker uses a pricing plan that reflects usage."
- Name
-
Official documentation · 21 Oct 2026
Page and navigation call the product Tinker; the page offers "Get started" and "Sign in".
- URL
-
Official documentation · 21 Oct 2026
HTTP 200 on 2026-10-21; Tinker's own page: "Frontier-grade infrastructure for training your own models."
- Deployment
-
Inferred 2026-09-08, not read off a page: every already-decided product of type 'api-service' in this catalogue includes "saas" in its deployment list, zero exceptions, per item 198's measured cross-tab. May be incomplete if this product also offers a self-hosted or private-cloud option; not wrong either way.
- Interfaces
-
Official documentation · 21 Oct 2026
CLI page: "tool manages sessions, checkpoints, and billing usage". The HTTP API is described at https://tinker-docs.thinkingmachines.ai/tinker/compatible-apis/openai/ (it exposes an OpenAI-compatible HTTP API); the SDK is already recorded.
- Modalities
-
Official documentation · 14 Oct 2026
Page: 'Add PNG or JPEG images and text to the same Inkling conversation' / 'Inkling natively supports audio input, which you can combine with other modalities such as text' / 'Generate tokens for interaction, evaluation, or RL actions'.
- Intended for
-
"a training API for researchers and developers"; case studies cite university research teams and enterprise companies building specialized AI applications
Also from Thinking Machines Lab
-
Tinker Serverless Inference API service
Beta pay-per-token hosted inference for Thinking Machines' open-weight Inkling and Inkling-Small models, priced separately from Tinker training; production use is by waitlist.
Open Thinking Machines Lab in the directory
Something wrong here? Send a correction — quote this product id: tinker.