A promise kept, on schedule
Start with what went right. The lab named a condition, a license and a rough date at launch, and met all three in seven days. The weights are the full model, not a distilled stand-in, and the license is Apache 2.0 as stated, subject to the company's acceptable use policy. Fine-tuning checkpoints built on Tinker during the closed week are portable now, which was the practical cost of the wait.
The release meets the announced schedule, with practical and evidentiary qualifications. The two qualifications that follow are about what "available" means in practice, and about the gate the release just passed through.
Downloadable is not the same as runnable
17B active parameters is the number that travels. The 17B figure counts active parameters per token, not total storage or runtime memory. The official card specifies at least 500GB of aggregated VRAM for its NVFP4 deployment. The MXFP8 alternative raises the floor to 800GB in exchange for a gain the card itself describes as very marginal and not meaningful.
The documented configuration targets multi-GPU hardware; the release does not supply a validated single-GPU deployment. For now, the practical routes are suitable multi-GPU hardware, or the hosted providers listed in the update, who will run it for everyone else.
The gap is already being worked on, and the first attempt shows its limits. A community conversion posted on Hugging Face on November 13 compresses the routed experts to about 110 GiB, sized for one 128GB DGX Spark. Its own card labels it experimental, reports 88.1% top-1 agreement in a teacher-forced fidelity test against the MXFP8 original, states that no task benchmark has been run, and states that its loader is still being ported, so the files do not load anywhere yet.
The gate produced a conclusion, not a record
The week's delay had a stated reason: external safety testing and red-teaming had to complete first. What the release publishes about that testing is a summary. The model card's safety section says evaluations covered everyday interaction, including multi-turn external red-teaming, and that CBRN and cyber capability were assessed through internal evaluations, external testing and refusal-suppressed variants. Thinking Machines concluded that the model did not present risk of material uplift beyond what is already available in the open-weight ecosystem, and that on loss-of-control measures it sits materially below frontier capabilities.
The card summarizes evaluation categories and conclusions but publishes no external scores, detailed reports or tester identities, limiting independent assessment. The November 12 update post does not mention the testing at all. A reader who wants to assess the judgment behind releasing the weights, rather than rely on the summary, has only that summary to work from. The conclusion may well be right. But a staged release process earns part of its credibility from the evidence it shows at each stage, and this stage shows the categories and the verdict, not the detailed working.
What changed for a decision
If your constraint was license or custody, it is gone: you can hold the weights, fine-tune them off Tinker, and ship products on them under Apache 2.0, subject to the acceptable use policy. If your constraint was hardware, the documented floor still applies until smaller faithful checkpoints exist and load. And if your evaluation process requires detailed external safety evidence before deploying an open model at this scale, the published summary may not be enough.
Sources
- [1] Thinking Machines, "Introducing Inkling 2 Small" with November 12, 2026 updateRead source
- [2] Thinking Machines, Inkling 2 Small model card, Hugging FaceRead source
- [3] Baekpica, Inkling-2-Small-VQ-GGUF model card, Hugging FaceRead source
- [4] ALFW, "Inkling 2 Small Is Here. Its Weights Are Not." (Nov 5, 2026)Read source