KINSMC
Inkling 2 Small launched through Tinker with closed weights. Thinking Machines Lab announced on November 5, 2026 that it intended to release the full weights under Apache 2.0 the following week, after external safety testing. Access to a hosted model and access to downloadable weights are different milestones. As checked on November 7, the official launch notice still describes the weight release as a future step.
For anyone deciding whether to try the model, build with it or run it on their own infrastructure, that distinction matters more than the word “open” in a headline. This guide separates the announced access routes from the deployment decisions that still require a released package and its accompanying instructions.
What is available, and what is still an announced plan?
| Access route | What the company describes |
|---|---|
| Tinker Playground | Hosted access for trying the model |
| Tinker developer platform | Sampling and fine-tuning access |
| Full downloadable weights | Planned for the following week, subject to completion of external testing |
| Planned weight licence | Apache 2.0; distinguish this announcement from the terms governing hosted use |
These are the manufacturer’s launch statements, rather than a claim that KINSMC has downloaded and tested a released weight package. The authoritative starting point is the Inkling 2 Small launch announcement.
What kind of model is Inkling 2 Small?
The official model card describes a multimodal model accepting text and images and producing text. It lists 594 billion total parameters, with 17 billion active, and a context window of up to one million tokens. The intended applications include coding assistants, tool-using agents and other developer-built AI systems.
The same card identifies Tinker as the distribution route and links its terms of use. It also distinguishes the Playground’s inference-only experience from developer access for inference and fine-tuning. These details are more useful for choosing an access route than treating every product carrying the model name as the same service.
Why hosted access is not the same as downloading weights
With a hosted service, you send a request to infrastructure operated by a provider and receive a response. Your application depends on that service’s interface, availability, account permissions and commercial terms. You can evaluate the model without necessarily possessing the files needed to operate it elsewhere.
Downloadable weights are the learned numerical values used by the model. Obtaining them can be one part of running a model independently, but it is not the whole deployment process. You also need a compatible implementation, suitable hardware, configuration and a clear understanding of the applicable licence.
For a small business, the useful comparison is therefore practical: do you need a quick hosted experiment, or do you have a reason and the resources to operate the system yourself? A future weight release does not automatically make self-hosting the cheapest or simplest option.
“17 billion active” does not mean a 17-billion-parameter download
In a mixture-of-experts architecture, only part of the model is used for a particular computation. The active-parameter figure describes that selected portion; the total-parameter figure describes a much larger collection of learned values. Those two numbers answer different questions.
A useful way to avoid an expensive mistake is to keep compute and storage questions separate. Ask how much work a request requires, then separately ask what must be stored and kept available during inference. Do not buy hardware on the assumption that the smaller number is the complete model footprint.
We are not giving a minimum GPU-memory requirement here. That would depend on the actual distribution, numerical format and supported serving setup. Wait for a documented configuration or a reproducible deployment report rather than converting a headline parameter count into a hardware promise.
How to approach a first evaluation
The official Inkling Cookbook guide identifies the model as thinkingmachines/Inkling-2-Small and recommends exploring reasoning effort and learning-rate settings for the relevant workflow. Use the maintained documentation for implementation details, because examples and supported settings can change.
For an initial business evaluation, our suggested process is deliberately small:
- Choose one real task. For example, extract fields from a document or propose a code change in a small project. Avoid judging the system only on an entertaining conversation.
- Prepare examples with known answers. Include ordinary cases, ambiguous cases and cases where the right answer is to acknowledge missing information.
- Record the configuration. Note the exact model, prompt, settings and evaluation date so a later comparison is meaningful.
- Score the useful outcome. Measure correct fields, passing checks or successful task completion, alongside response time and cost.
- Review failures. A fluent explanation of an incorrect result is still an incorrect result.
This is an evaluation framework, not a report of tests we have conducted on Inkling 2 Small. Keeping that distinction explicit allows readers to reproduce their own assessment instead of mistaking a launch explainer for a benchmark review.
What to check when the weights are released
Start from the company’s own announcement and follow its official repository links. Confirm the exact model name and version, read the licence packaged with the files and check whether the instructions cover your intended environment. A repository for an earlier Inkling model is not evidence that the second-generation package is available.
Then examine the practical requirements: file sizes, supported numerical formats, serving software and any limitations described by the maintainer. Keep hosted-service terms and downloadable-file licensing separate in your records. The fact that two routes use the same model name does not make their operational obligations identical.
Frequently asked questions
Can I try Inkling 2 Small before downloading it?
The company describes hosted access through Tinker. Check the platform’s current account requirements and terms before using it.
Does a planned Apache 2.0 release mean the files are already downloadable?
No. A licensing plan and a completed distribution are different events. Look for the actual official release, not only a statement of intent.
Has KINSMC independently tested its performance?
No. This article explains the official release status and offers a practical evaluation checklist. It does not present hands-on benchmark results.
Source check: November 7, 2026. Prepared with AI assistance using the official sources linked above. Featured image: AI-generated conceptual illustration, not a product screenshot.