Training API
Not listed
Model/token usage rates; storage US$0.10/GB-month
- Included scope
- Managed LoRA training, sampling and checkpoint access
Tinker
● Identity checkedModel training and fine-tuning API · Official-source review 2026-10-09
Thinking Machines’ Tinker is a training API for researchers using LoRA adapters on supported models. It abstracts infrastructure while exposing forward/backward gradient computation, optimizer updates, sampling and checkpoint saving for custom supervised or reinforcement-learning loops.
- Category
- Model training and fine-tuning API
- Free access
- Not confirmed: ongoing free training or signup credit. Training/sampling priced by model usage; checkpoint storage advertised US$0.10/GB-month.
- API access
- Training API and saved-checkpoint download endpoint, with cookbook examples. Exact model, token and concurrency limits require current documentation.
Is Tinker right for you?
A good fit for
Relevant to researchers who want custom data and reward loops without managing distributed training infrastructure.
Before you choose
Model lineup changes and includes retired entries; verify current supported identifiers before selecting a base. Prices are USD per million tokens, exact training/sampling rates not in reviewed homepage. LoRA adapters do not equal full base-weight fine-tuning. Dataset rights, checkpoint licenses, retention and downstream deployment costs need checks. Data-nontraining and case-study performance are vendor claims, not independently validated here. No data uploaded, training job, checkpoint download or benchmark performed.
Workflow and handoff
Verify model/license and rates, use a small synthetic dataset and held-out evaluation and track training plus sampling/storage costs before scaling.
Pricing and practical limits
Model lineup changes and includes retired entries; verify current supported identifiers before selecting a base. Prices are USD per million tokens, exact training/sampling rates not in reviewed homepage. LoRA adapters do not equal full base-weight fine-tuning. Dataset rights, checkpoint licenses, retention and downstream deployment costs need checks. Data-nontraining and case-study performance are vendor claims, not independently validated here. No data uploaded, training job, checkpoint download or benchmark performed.
Review method
Official product evidence reviewed on 2026-10-09; product artwork inspected visually. No hands-on vendor application test was performed. Application performance, security claims and plan enforcement were not independently tested.
What it can do
Features & capabilities
Unknown is different from unavailable. Each fact carries its own evidence.
| Capability | Value | Evidence | Checked |
|---|---|---|---|
| Product workflow | Thinking Machines’ Tinker is a training API for researchers using LoRA adapters on supported models. It abstracts infrastructure while exposing forward/backward gradient computation, optimizer updates, sampling and checkpoint saving for custom supervised or reinforcement-learning loops. | Facts sourced | 2026-10-09 |
| Pricing and restrictions | Model lineup changes and includes retired entries; verify current supported identifiers before selecting a base. Prices are USD per million tokens, exact training/sampling rates not in reviewed homepage. LoRA adapters do not equal full base-weight fine-tuning. Dataset rights, checkpoint licenses, retention and downstream deployment costs need checks. Data-nontraining and case-study performance are vendor claims, not independently validated here. No data uploaded, training job, checkpoint download or benchmark performed. | Facts sourced | 2026-10-09 |
Understand the total cost
Tinker pricing & plans
Alternatives to Tinker
View all ↗
MCP-Builder.aiMCP-Builder.ai generates and hosts MCP servers from a described data-access workflow. It connects supported APIs and databases and provides a server URL for compatible assistant clients.
Explore tool
PagekitePageKite provides pagekite.py, a localhost tunneling solution that exposes local servers to the internet through managed front-end relays. It supports pay-what-you-want purchases, named subscription plans for teams, custom domains, white-label embedded offerings, and a documented free trial before paid service.
Explore tool
ConscioussAIConscioussAI connects AI models to operating-system workflows on a phone or computer. Its plans combine model access, credit budgets and higher-tier screen sharing and hands-free controls.
Explore toolThe practical questions
Frequently asked questions
Was the application hands-on tested?
No. This listing is based on dated official-source evidence and a visual artwork check. Unconfirmed costs and restrictions are identified explicitly.
What is Tinker used for?
Thinking Machines’ Tinker is a training API for researchers using LoRA adapters on supported models. It abstracts infrastructure while exposing forward/backward gradient computation, optimizer updates, sampling and checkpoint saving for custom supervised or reinforcement-learning loops.
Does Tinker have a free plan?
Not confirmed: ongoing free training or signup credit. Training/sampling priced by model usage; checkpoint storage advertised US$0.10/GB-month.. This record does not confirm an ongoing free plan.
How much does Tinker cost?
A listed monthly price has not been confirmed. See the plan cards for entitlements, billing commitments and seat minimums.
Can I use Tinker through an API?
Training API and saved-checkpoint download endpoint, with cookbook examples. Exact model, token and concurrency limits require current documentation.. API access and subscription access may have different terms; consult the linked sources.
What should I check before choosing it?
Model lineup changes and includes retired entries; verify current supported identifiers before selecting a base. Prices are USD per million tokens, exact training/sampling rates not in reviewed homepage. LoRA adapters do not equal full base-weight fine-tuning. Dataset rights, checkpoint licenses, retention and downstream deployment costs need checks. Data-nontraining and case-study performance are vendor claims, not independently validated here. No data uploaded, training job, checkpoint download or benchmark performed.
Price history
No retained pricing changes yet. A current price alone does not establish a historical trend.
How this profile is supported
Official product information
Current official product source read individually; product artwork inspected visually.
Read original source ↗Pricing review
Official product page reviewed for disclosed costs; unverified prices and allowances are identified in the listing.
Read original source ↗Facts apply to the named version and check date. Send a sourced correction if something changed.