Thinking Machines Says Inkling 2 Small Weights Arrive Under Apache 2.0 "Next Week." Plan the Test, Not the Migration.
On November 5, Thinking Machines released Inkling 2 Small, a mixture-of-experts model with 594 billion total parameters and 17 billion active, available for now only through its Tinker platform with closed weights. The company says it intends to publish the full weights under Apache 2.0 "next week," after external safety testing and red-teaming. Relative to the November 5 post, that points at the week of November 9. A stated intention is not a shipped license, so the useful move today is preparing the evaluation, not the migration.
What is actually announced
From the model page: 594B total and 17B active parameters, a 292M-parameter vision encoder, and a benchmark table that includes an Artificial Analysis Intelligence Index score of 31, SWE-Bench Pro Hard at 86.3% and DeepSWE at 72.1%. Those are the company's own numbers, run by the company. Treat them as a claim to reproduce, not a ranking.
The safety section is candid. The team says CBRN scores are below the open-weight frontier, cyber capability is well above the earlier Inkling Small but behind the open-weight frontier, and early checkpoints sometimes fabricated results, ignored instructions or concealed actions before alignment training. I could not find a Tinker price on the page, so I cannot give you a cost per task.
Why "open weights" may not mean "I can run it"
Seventeen billion active parameters sounds small. It is not the number that decides your hardware. In a mixture-of-experts model, all 594B parameters have to be in memory even though only some are used per token. At 8-bit precision that is on the order of 600 GB of weights before context cache. That is a multi-GPU server, not a laptop and not a single consumer card. This is my arithmetic from the parameter count, not a figure from the vendor.
So for most solo builders, the realistic path is a hosted endpoint from whoever serves it first, which means the Apache 2.0 license matters mainly because it lets several providers compete on price.
The plan I would have ready
- Write down the go/no-go before the weights drop. Pick 20 real tasks from your own product, not public benchmarks, and a pass threshold.
- Run the same 20 tasks against your current model and record cost and failures. That is the baseline.
- When a provider lists Inkling 2 Small, run the 20 tasks once. If it beats your threshold at lower cost, run it again with your actual traffic shape.
- Confirm the license text when it is published. "Apache 2.0" is the stated plan; read the published file and any usage policy that ships beside it.
- Keep your model call behind one thin function, so swapping is a config change.
The honest take
The counter-argument is simple: weights have been promised "soon" before and arrived late, narrower or with extra conditions. If that happens here, the 20-task harness costs you an afternoon and still pays off against your next model. The benchmark table might also hold up fine; I have not run any of it. What I would not do is rewrite a working stack on the strength of a press page and a date that says "next week."
Author
Lukas
@lukcombinator