Thinking Machines announced Inkling 2 Small on 5 November. Tinker offers inference and fine-tuning now, but the weights remain closed. The company intends an Apache 2.0 weight release the following week, after external safety testing and red-teaming. Launch announcement.
For a team choosing a model, availability and suitability are separate decisions. You can prepare an evaluation without committing a production workflow to the model.
Read the model label carefully
The mixture-of-experts model has 594 billion total parameters and 17 billion active. The vendor flags verbosity and imperfect tool choices among its limitations. Architecture and limitations. The word “Small” is not a consumer-hardware specification.
Artificial Analysis lists text/image input, text output and a 1M-token model context, while its provider listing reports 262k. Check the selected endpoint’s actual limit rather than treating model capacity as a service guarantee. Model profile; provider profile.
Compare a finished task, not just a token
Independent evaluator Artificial Analysis reports 31 on its Intelligence Index v5.0.1. Its tests averaged roughly 150,000 output tokens and $1.18 per task, versus 37,000 and $0.20 for the previous Inkling Small. It lists API rates of $0.30 per million input tokens and $1.00 per million output tokens. All dollar amounts are USD. These are dated evaluation figures, not a quote for your workload. Independent assessment, 5 November.
A benchmark score can help select a candidate, but cannot define acceptance for your application. A longer answer is useful only if it completes the job well enough to justify the time, review and processing it requires.
Make the comparison repeatable
- Choose representative work: use a small set of non-sensitive examples with known acceptance criteria, including cases the system should decline or escalate.
- Keep the setup comparable: record the model version, reasoning setting, tools and retry policy. Changing several at once obscures the comparison.
- Measure the whole job: record success, elapsed time, tokens, retries, tool expenses and human corrections. Agree a spending ceiling before running anything.
- Review deployment separately: check current prices, endpoint limits and data-handling terms. For self-hosting, wait for the actual weights, license and deployment documentation.
These are proposed evaluation steps, not hands-on results. Treat the launch as a reason to test a candidate against your requirements, not as evidence that it is universally better.