Thinking Machines Lab’s Tinker: Service and Business Model

Thinking Machines Lab’s Tinker is a generally available managed service for fine-tuning open-weight models, not a packaged model or no-code trainer. Customers control their data and post-training logic; Tinker provides access to supported models and runs the distributed accelerator infrastructure behind each job.

The business model follows that division of labor. Thinking Machines Lab handles remote execution, scheduling, resource allocation and recovery from infrastructure failures, then charges for the compute and storage consumed. The customer remains responsible for examples, rewards, evaluations and algorithm choices.

What Tinker provides

Tinker turns post-training into a set of programmable service calls. A developer can keep orchestration code on a CPU machine while requesting remote forward-and-backward computation, optimizer steps, sampling and checkpoint operations. This preserves control over the experiment without requiring the customer to assemble and operate the underlying training cluster.

In the official Tinker launch description, Thinking Machines Lab says the service handles distributed training, scheduling, resource allocation and failure recovery. It exposes low-level operations including forward_backward and sample, and uses LoRA so multiple training runs can share a compute pool rather than modifying every model parameter.

That operational layer is the core product. The service removes cluster management from the customer’s workload, but it does not decide what the model should learn or whether an experiment has succeeded.

What remains under customer control

The customer supplies training examples or an interactive environment, formats model inputs, defines the loss or reward signal, chooses hyperparameters and decides when to sample, evaluate and update weights. For supervised fine-tuning, that includes preparing examples and loss masks; for reinforcement learning, it can include generating rollouts, scoring them and calculating advantages.

This makes Tinker a service for AI researchers, model-training engineers and technically capable startups that want low-level experimental control without running distributed infrastructure. A team looking for a dataset upload form and an automatically selected recipe would need a higher-level layer.

The product map

Managed infrastructure: remote accelerator capacity, distributed execution, scheduling, resource allocation and failure recovery. This is the principal paid service.
Training interface: programmable primitives for gradients, optimizer steps, sampling, log probabilities and weight management.
Post-training methods: supervised fine-tuning, reinforcement learning, preference optimization and distillation can be assembled with those primitives. They are workflows, not separate hosted products.
Model access: customers select from the models supported by the platform; availability, context limits and rates depend on the model.
Open-source cookbook: recipes and abstractions demonstrate training patterns and reduce implementation work, while execution on Tinker’s managed infrastructure remains billable.

The cookbook and the commercial service therefore solve different problems. The code helps a team express an experiment; Tinker supplies the remote model access, compute and operational machinery needed to execute it at scale. Open-source recipes do not make the provider’s clusters or serving capacity open source.

How usage becomes revenue

A Tinker reinforcement-learning run consumes prefill, sampling, training and checkpoint storage as the loop progresses.

Tinker meters activity instead of charging one flat fine-tuning fee. Its models and pricing documentation publishes model-specific per-million-token rates for prefill, sampling and training, defines training as the forward-and-backward work that produces gradients, and lists checkpoint storage at $0.10 per gigabyte per month. Cached prefill receives an 80% discount, while some model rates may carry temporary discounts.

The experiment’s shape therefore drives the bill. Longer inputs consume more prefill; repeated rollout generation increases sampling; larger batches or additional update cycles increase training usage; and retained checkpoints consume storage. Model choice matters because rates and context limits vary across the supported catalogue.

A reinforcement-learning loop may incur charges at several stages as it processes prompts, generates rollouts and performs updates. The commercially relevant unit is consequently the whole training workflow, not only the advertised training-token rate.

From private beta to general availability

Tinker launched in private beta on October 1, 2025. Thinking Machines Lab’s December 12 general-availability update removed the waitlist, introduced OpenAI API-compatible sampling scaffolding and added image inputs for two Qwen3-VL models available at that time. The compatibility layer allowed software using that API format to sample from a Tinker model by its model path, including during training.

Image input expanded the service beyond text-only workflows: encoded text could be interleaved with image bytes for supervised or reinforcement-learning runs. The two models named in that release have since appeared on Tinker’s retired-model list, illustrating that model access is a managed, changing catalogue rather than permanent inventory.

What the customer is buying

Self-hosting comparable work requires accelerator capacity, distributed-training software, model loading, scheduling, checkpoint storage, monitoring and recovery when workers fail. Tinker absorbs that operational burden while retaining a relatively low-level programming surface for the customer.

The trade-off is precise: infrastructure control moves to Thinking Machines Lab, while experiment design stays with the customer. Tinker sells managed execution to teams that already know what training loop they want to build—not automatic model customization, but the compute access and operations needed to run that loop without becoming a cluster operator.

Thinking Machines Lab’s Tinker: What Its Fine-Tuning Service Sells