You can try Inkling 2 Small in your browser through Tinker Playground. Thinking Machines Lab also offers it through the Tinker API for generating responses and fine-tuning. The model weights, the files you would need to run it on your own hardware, are not yet available. The company said on Nov. 5 that it plans to release them under Apache 2.0 after external safety testing. Launch announcement
Try one real task on Tinker
Open Tinker Playground, sign in and choose Inkling 2 Small. The older Inkling-Small is a separate model. Start with a task you can check yourself. If you want help tracing an error in a chart, use a chart you’re allowed to upload and compare the answer with the underlying numbers. The model accepts images as well as text, and returns text. Try a follow-up asking it to point to the values it used so you can catch a misread axis or unit. Inkling 2 Small model card
What Tinker charges
Tinker lists the following prices per million tokens. Both options charge $0.03 for a million cached input tokens when the cache discount applies. Tinker models and pricing
| Tinker model option | Input | Output | Training |
|---|---|---|---|
| 64K context | $0.30 | $1.00 | $1.50 |
| 256K context | $0.30 | $1.00 | $3.00 |
A request with 10,000 uncached input tokens and 2,000 output tokens would cost about half a cent at those rates. Repeated requests and longer outputs increase that total. Tinker lists checkpoint storage separately at $0.10 per GB per month. The company describes support for up to one million tokens in its launch announcement. The two priced options listed by Tinker are 64K and 256K. Use the limits for the option you select when planning a job. Launch announcement · Tinker pricing
Use the 64K option if your prompts and replies fit within that context. The 256K option can accommodate more material in a single request, and its training rate is twice as high. The listed input and output rates are the same for both. Tinker pricing
Using the Tinker API
If you’re calling the model from code, Tinker’s API quickstart explains how to set up access. The standard model ID in the pricing table is thinkingmachines/Inkling-2-Small. Sampling is the step that generates a response. Fine-tuning creates a checkpoint from your training examples. Try sampling first if you’re still changing your prompts or haven’t set aside examples for evaluating a tuned model.
Why 17 billion active parameters do not tell you how much memory you need
The company lists 594 billion total parameters and 17 billion active parameters. Inkling 2 Small is a mixture-of-experts model: it uses only a portion of the model for a given token. A machine still needs a way to store and access the other weights as different experts are selected. The active count therefore cannot tell you whether the full model will fit on your machine. Architecture details
As a sizing illustration, storing 594 billion numbers at two bytes each takes roughly 1.19 TB before runtime overhead. At a theoretical four bits per parameter, the raw weights alone would take about 297 GB. Those are arithmetic estimates, not measured download sizes or hardware requirements. Actual files can include extra data, and a particular runtime might keep some weights in system RAM or distribute them across several GPUs. The format and runtime matter as much as the number on the spec sheet.
Before choosing hardware, wait for the release repository to list the available file formats and for the runtime instructions to specify supported devices. The official files and accompanying license will determine what you can run and redistribute. Thinking Machines Lab release announcement
