Tinker…The product
Three inputs. One ready-to-run training loop.
Choose a verifier, a task, and a base model. Download a ready-to-run training loop.
How this works
Choose a verifier, a task, and a base model. The compiler writes the loop. Ten example files below already compiled.
Training signal
How will you score an output?
Verifiable outcome, preference pairs, or demonstrations. That chooses the loss and the data shape. A verifier returns a reward. That is not yet a learned reward model.
Task
Which behavior should improve?
Tool agents, calibrated forecasting, or formal reasoning. Same training signal, different traces. Prompts and checkers change. The algorithm does not.
Base model
Which open weights get the LoRA?
Dense if every weight should be live. MoE if you want active-parameter thrift. Tinker hosts the GPUs. The compiler writes one BASE_MODEL string.
Pipeline
Training signal × task × base model → sample, forward_backward, optim_step, save_state.
sample → forward_backward → optim_step → save_state. Those four calls are the file you download.
10 compiled walks
Every path below produced a Python file with sample, forward_backward, optim_step, and save_state. Changing the task or base model never dropped a primitive. Changing the training signal is what changes the loss.
No.01 Signal02 Task03 Base modelYou get
01Verifiable outcomeSpecialized Tool AgentsNemotron-3-NanoRunnable environment loop
02Verifiable outcomeFormal Reasoning EnginesDeepSeek-V3.1Runnable environment loop
03Verifiable outcomeCalibrated ForecastingQwen3.6-35B-A3BRunnable environment loop
04Preference pairsSpecialized Tool AgentsQwen3-8BRunnable preference loop
05Preference pairsCalibrated ForecastingDeepSeek-V3.1Runnable preference loop
06Preference pairsFormal Reasoning EnginesQwen3.5-9BRunnable preference loop
07DemonstrationsSpecialized Tool AgentsNemotron-3-NanoRunnable distillation loop
08DemonstrationsCalibrated ForecastingQwen3.5-4BRunnable distillation loop
09Verifiable outcomeSpecialized Tool AgentsQwen3-8BRunnable environment loop
10DemonstrationsFormal Reasoning EnginesQwen3.6-35B-A3BRunnable distillation loop
Tinker · Thinking Machines Lab
The key stays on your machine. The script stays yours.
TinkerTraining API · four primitivesReinforcement compiles. The file calls sample, forward_backward, optim_step, and save_state. It reads TINKER_API_KEY from the environment. The key is never written into the download. Training, pricing, and GPU scheduling are Tinker.
Site map
Build, How this works, Docs, Cookbook, Models, Example compiles, Blog, About, and the HackerNoon AI & ML Newsletter. Terms and Privacy stay in the footer.
Now
- BuildChoose a verifier, a task, and a base model. This is the product.
- How this worksWalkthrough. Ten files this compiler already wrote.
- DocsCompiler contract, knobs, loss types, and how the downloaded script talks to Tinker.
- CookbookNamed recipes. Same four primitives, different training signals.
- ModelsCurated Tinker bases, Inkling first. Live catalog only when this host can reach Tinker.
- Example compilesTen compiled files. Not a stored run history. Your key, your machine.
- BlogWriting on HackerNoon under /u/reinforcement.
- AboutWho built Reinforcement.tech and why the training signal comes first.
- HackerNoon AI & ML NewsletterHackerNoon AI & ML — https://hackernoon.com/c/ai
- TermsBarebones terms for an alpha compiler.
- PrivacyNo account. No key in the download. What we do and do not keep.
Next
- StatusA dedicated health page. The titlebar chip already pings get_server_capabilities.
- ChangelogWhat shipped. What broke. What the compiler now writes.
- Custom verifierBring a verifier that is not on the desk yet. The rest of the pipeline stays the same.